fix(bench): chat scenarios skip decision models (#341)

helexa-bench runs its chat scenarios against every warm model except
image generators. Laya (capabilities ["decision"]) is now resident on
quadbrat and answers /v1/chat/completions with 422 wrong_modality, so
every sweep recorded failed chat runs against it and logged a rejection
on the neuron. The default scenario gate now excludes every non-chat
modality -- image and decision -- through one serves_chat() check.

Mutation-checked: restoring the image-only gate fails the new test.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01VtaCXFAY23HmwX5Cyc6jhC

.gitea/workflows/build-prerelease.yml

@@ -525,7 +525,14 @@ jobs:
             cuda_home: /usr/local/cuda-13.0
             build_jobs: 8
             nvcc_threads: 4
-            cargo_features: "cuda cudnn"
+            # flash-attn on ampere (#337): Laya's encoder runs every
+            # attention through flash_attn_varlen_windowed, matching the
+            # PyTorch reference on a 3060 (36 vs 39 ms at 512 tokens) where
+            # the eager path is 1.4-1.7x slower. This also moves the
+            # flavour's other flash-capable paths onto flash -- qwen3_5
+            # full attention and Z-Image -- which validate-neuron.sh and
+            # validate-image.sh gate. NEURON_FLASH_ATTN=0 rolls back.
+            cargo_features: "cuda cudnn flash-attn"
           - flavour: ada
             compute_cap: "89"
             runner: cuda-13.0
@@ -540,8 +547,8 @@ jobs:
             build_jobs: 8
             nvcc_threads: 4
             # flash-attn on blackwell first (#95): beast carries the
-            # agentic prefill pain; ada/ampere follow once the win is
-            # measured. NEURON_FLASH_ATTN=0 is the runtime rollback.
+            # agentic prefill pain. NEURON_FLASH_ATTN=0 is the runtime
+            # rollback.
             cargo_features: "cuda cudnn flash-attn"
     runs-on: ${{ matrix.runner }}
     env:

.gitea/workflows/deploy.yml

@@ -244,10 +244,14 @@ jobs:
             flavour: blackwell
             load_timeout: 900
             max_prompt_tokens: 131072
-          - host: benjy.hanzalova.internal
-            flavour: ada
-            load_timeout: 300
-            max_prompt_tokens: 16384
+          # benjy is out of the matrix while its boot NVMe is replaced
+          # (dead 2026-09-29): every deploy failed at SSH init against a
+          # host that is not there. Restore these four lines when it is
+          # back on the mesh.
+          # - host: benjy.hanzalova.internal
+          #   flavour: ada
+          #   load_timeout: 300
+          #   max_prompt_tokens: 16384
           - host: quadbrat.hanzalova.internal
             flavour: ampere
             load_timeout: 300

Cargo.lock

@@ -3149,6 +3149,8 @@ dependencies = [
  "image",
  "libc",
  "memmap2",
+ "metrics",
+ "metrics-exporter-prometheus",
  "minijinja",
  "minijinja-contrib",
  "rayon",
@@ -3164,6 +3166,8 @@ dependencies = [
  "toml",
  "tracing",
  "tracing-subscriber",
+ "unicode-normalization",
+ "unicode-properties",
 ]
 
 [[package]]

asset/monitoring/README.md

@@ -6,10 +6,18 @@ bench UI shows.
 
 ## Topology
 
-- **cortex** runs on `hanzalova.internal` (10.6.0.46) and is the **only**
-  Prometheus target — it exposes every `cortex_*` metric on `:31314`, already
-  labelled by `{node,model}` / `{node,device}` from its neuron poller. neuron
-  has no `/metrics` endpoint.
+- **cortex** runs on `hanzalova.internal` (10.6.0.46) and exposes every
+  `cortex_*` metric on `:31314`, already labelled by `{node,model}` /
+  `{node,device}` from its neuron poller. This is the fleet-shaped view:
+  who is loaded where, how busy, how much VRAM.
+- **neuron** additionally serves `/metrics` on its own API port
+  (`:13131`) on each GPU host. This is the request-shaped view — where a
+  request's time actually went, as histograms rather than polled gauges.
+  It exists because cortex's numbers are a ~10s poll of `/health`
+  republished as gauges, which cannot see inside a one-second forward and
+  carries no distribution. Scrape job: `prometheus-neuron.scrape.yml`.
+  **No firewalld rule needed** — 13131 is already open for the control
+  plane (`data/neuron-firewalld.xml`).
 - **Prometheus + Grafana** run on `golgafrinchans.kosherinata.internal` as
   podman quadlets. Host ports (from the registered band — 9090 is Cockpit,
   3000 was avoided): **Prometheus `:26559`** (container config bind-mounted at

asset/monitoring/prometheus-cortex.scrape.yml

@@ -4,9 +4,12 @@
 # Cockpit on that box):  curl -X POST http://localhost:26559/-/reload
 #
 # cortex runs on hanzalova.internal (10.6.0.46) and exposes ALL fleet metrics
-# on :31314 — it is the only Prometheus target in the stack (neuron has no
-# /metrics; cortex's poller already relabels everything by {node,model} /
-# {node,device}). Reaching it cross-DC (kosherinata -> hanzalova) requires the
+# on :31314; cortex's poller already relabels everything by {node,model} /
+# {node,device}. It is no longer the only target — neuron now serves its own
+# /metrics on :13131 with the per-request phase histograms cortex's ~10s
+# gauge poll cannot express (see prometheus-neuron.scrape.yml); the two are
+# complementary, not redundant.
+# Reaching cortex cross-DC (kosherinata -> hanzalova) requires the
 # firewalld rule in asset/monitoring/firewalld-cortex-metrics.sh, which opens
 # :31314 to this host's mesh IP (10.3.101.4) only.
 #

asset/monitoring/prometheus-neuron.scrape.yml

@@ -0,0 +1,40 @@
+# helexa neuron scrape job — append into the `scrape_configs:` list of
+# /etc/prometheus/prometheus.yml on golgafrinchans.kosherinata.internal,
+# then reload:  curl -X POST http://localhost:26559/-/reload
+#
+# WHY A SECOND TARGET, when cortex already publishes fleet metrics.
+#
+# cortex's numbers come from polling each neuron's /health every ~10s and
+# republishing it as gauges. That is the right shape for routing decisions
+# and the wrong shape for asking where a request's time went: an EMA
+# sampled at 10s cannot see inside a one-second forward, and gauges carry
+# no distribution. The finest split cortex can express is prefill vs
+# decode.
+#
+# neuron serves /metrics on its own API port (13131) — the port is already
+# open in data/neuron-firewalld.xml for the cortex control plane, so unlike
+# the cortex scrape this needs NO new firewalld rule.
+#
+# What is here that cortex cannot give you:
+#   neuron_decode_phase_seconds{model,phase=forward|sample|emit|residual}
+#     Decode time by phase, summed per request, as a real histogram.
+#     The `_sum` series is the point — ratio a phase against the whole to
+#     see where decode goes:
+#       sum(rate(neuron_decode_phase_seconds_sum{phase="forward"}[5m]))
+#         / sum(rate(neuron_decode_phase_seconds_sum[5m]))
+#     `residual` is decode_ms minus the three measured phases, emitted as
+#     its own series so unbracketed time is visible rather than divided
+#     silently among the phases that were measured.
+#   neuron_prefill_seconds / neuron_decode_seconds  (histograms)
+#   neuron_completion_tokens_total  (the per-token denominator)
+#
+# A model whose serving path does not measure phases emits NO phase series
+# at all rather than zeros — absent means "nobody looked", not "free".
+- job_name: neuron
+  static_configs:
+    - targets:
+        - 'beast.hanzalova.internal:13131'
+        - 'benjy.hanzalova.internal:13131'
+        - 'quadbrat.hanzalova.internal:13131'
+      labels:
+        fleet: hanzalova

asset/neuron/quadbrat.toml

@@ -7,9 +7,10 @@
 # ─────────────────────────────────────────────────────────────────────
 # neuron.toml for quadbrat.hanzalova.internal
 #
-# 1x RTX 3060 (12 GB) — small / quantised tier. Pre-warms Qwen3-1.7B
-# (bf16, ~4 GB), leaving ~7 GB for KV cache so long contexts on a small
-# model still have plenty of room.
+# 1x RTX 3060 (12 GB). Resident: Laya, the System-One decision model
+# (#333, ~2.3 GB bf16 for its three checkpoints). It replaced the
+# Qwen3-1.7B pre-warm (operator decision 2026-09-29, #341); the 1.7B is
+# still cold-loadable beside it for helexa/small.
 #
 # Synced to /etc/neuron/neuron.toml by script/infra-setup.sh.
 
@@ -21,10 +22,21 @@ name = "candle"
 [harness.candle]
 
 [[default_models]]
-model_id = "Qwen/Qwen3-1.7B"
+model_id = "convaiinnovations/laya"
 harness = "candle"
 devices = [0]
 
+# A decision request is one forward of tens of milliseconds (36 ms for
+# one question over a 512-token state, 162 ms for five), so a one-slot
+# queue of 8 rejects ordinary bursts. Admit
+# several at once -- the device worker still runs forwards one at a
+# time -- and let a deep, short-deadline queue absorb the rest. Applies
+# to every model on this host, including a cold-loaded helexa/small.
+[harness.candle.admission]
+max_in_flight = 4
+max_queue_depth = 64
+max_wait_secs = 10
+
 # Text-to-image (#204): the 12 GB tier serves the q8_0-quantized DiT
 # (8.7 GB resident incl. VAE). Ceiling is 768² — naive f32 attention
 # transients OOM 1024² beside the resident model; lifts to 1024² if

crates/cortex-core/src/catalogue.rs

@@ -107,6 +107,27 @@ pub const DEFAULT_RESIDENCY_PRIORITY: u32 = 100;
 /// grant on its own.
 pub const PINNED_RESIDENCY_PRIORITY: u32 = 1000;
 
+/// The alias a `/v1/systemone` decision request routes to when its
+/// `model` is absent or names nothing this fleet serves.
+///
+/// Decision clients written against TypeSafe's Jev API send a `model`
+/// naming *their* service (e.g. `jev-1`), and Laya's own server treats
+/// any unrecognised value as "let the router choose". Every hop of the
+/// serving chain applies the same rule, scoped to the decision endpoint
+/// only — a chat request for an unknown model is still a 404. The
+/// operator points the alias at a decision model in `[aliases]`.
+pub const DECISION_DEFAULT_ALIAS: &str = "helexa/one";
+
+/// Header carrying the concrete decision model a request was routed to.
+///
+/// A decision request's `model` field can name a checkpoint *inside* a
+/// served model family (`multilingual`, `typed-decisions`) or a foreign
+/// service, so the gateway forwards the body untouched and states the
+/// routing outcome here instead of rewriting `model`. The serving node
+/// uses it to pick the model and still reads `model` for a checkpoint
+/// pin.
+pub const DECISION_MODEL_HEADER: &str = "x-helexa-decision-model";
+
 /// The full model catalogue.
 #[derive(Debug, Clone, Serialize, Deserialize, Default)]
 pub struct ModelCatalogue {

crates/cortex-core/src/lib.rs

@@ -2,6 +2,7 @@ pub mod anthropic;
 pub mod build_info;
 pub mod catalogue;
 pub mod config;
+pub mod decisions;
 pub mod discovery;
 pub mod entitlements;
 pub mod error_envelope;

crates/cortex-core/src/openai.rs

@@ -210,6 +210,19 @@ pub struct HelexaTiming {
     pub prefill_ms: u64,
     pub decode_ms: u64,
     pub prefill_tokens: u64,
+    /// Decode-loop phase split, when the serving path measured it.
+    /// Absent (rather than zero) on paths that do not, so a consumer
+    /// can tell "not measured" from "measured as free". Together these
+    /// should roughly account for `decode_ms`; the residual is time in
+    /// the loop that no bracket covers.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub forward_ms: Option<u64>,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub sample_ms: Option<u64>,
+    /// Detokenise + hand to the consumer. Includes bounded-channel
+    /// backpressure, so a slow reader is charged here.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub emit_ms: Option<u64>,
 }
 
 /// Sub-counts of `Usage::completion_tokens`.

crates/cortex-gateway/src/handlers.rs

@@ -24,6 +24,7 @@ pub fn api_routes() -> Router<Arc<CortexState>> {
         .route("/v1/completions", post(completions))
         .route("/v1/responses", post(responses))
         .route("/v1/images/generations", post(images_generations))
+        .route("/v1/systemone", post(systemone))
         .route("/v1/models", get(list_models))
         .route("/v1/messages", post(anthropic_messages))
         .route("/health", get(health))
@@ -349,6 +350,192 @@ async fn images_generations(
         })
 }
 
+/// `POST /v1/systemone` — typed decisions (TypeSafe Jev wire protocol)
+/// from a decision model such as Laya (#338).
+///
+/// Routing differs from the other inference paths in one way: a
+/// missing or unknown `model` resolves to
+/// [`DECISION_DEFAULT_ALIAS`](cortex_core::catalogue::DECISION_DEFAULT_ALIAS)
+/// instead of 404ing, because Jev clients name their own service there.
+/// A `model` that *does* resolve — an alias, or a model this fleet
+/// knows — routes as usual, so asking a text model for a decision
+/// reaches neuron and gets its `wrong_modality` answer rather than a
+/// silent substitution.
+///
+/// The body is forwarded untouched (its `model` may pin a checkpoint
+/// inside the served family); the routed model travels in
+/// [`DECISION_MODEL_HEADER`](cortex_core::catalogue::DECISION_MODEL_HEADER).
+///
+/// Metered like Jev: input tokens only. The reservation is the exact
+/// upper bound — questions × the sequence cap — and settles to
+/// `usage.input_tokens` from the response.
+async fn systemone(
+    State(fleet): State<Arc<CortexState>>,
+    headers: HeaderMap,
+    body: Bytes,
+) -> Response {
+    use cortex_core::catalogue::{DECISION_DEFAULT_ALIAS, DECISION_MODEL_HEADER};
+
+    log_inbound("systemone", "/v1/systemone", &body);
+    let requested = extract_model(&body);
+    let route = match requested.as_deref() {
+        Some(model) => match router::resolve(&fleet, model).await {
+            Err(router::RouteError::ModelNotFound(_)) => {
+                tracing::debug!(
+                    requested = model,
+                    default = DECISION_DEFAULT_ALIAS,
+                    "unknown decision model; using the default"
+                );
+                router::resolve(&fleet, DECISION_DEFAULT_ALIAS).await
+            }
+            other => other,
+        },
+        None => router::resolve(&fleet, DECISION_DEFAULT_ALIAS).await,
+    };
+    let route = match route {
+        Ok(r) => r,
+        Err(e) => {
+            tracing::warn!(
+                handler = "systemone",
+                requested = requested.as_deref().unwrap_or("<none>"),
+                error = %e,
+                "route resolve failed"
+            );
+            return route_error_response(&e);
+        }
+    };
+
+    touch_model(&fleet, &route.node_name, &route.resolved_model_id).await;
+
+    let labels = [
+        ("model", route.resolved_model_id.clone()),
+        ("node", route.node_name.clone()),
+    ];
+    metrics::counter!("cortex_requests_total", &labels).increment(1);
+    metrics::counter!("cortex_decision_requests_total", &labels).increment(1);
+    if route.cold_start {
+        metrics::counter!("cortex_cold_starts_total", &labels).increment(1);
+    }
+
+    let principal = crate::metering::principal_from_headers(&headers);
+    let bound = crate::metering::decision_reservation_tokens(&body);
+    let guard = match &principal {
+        Some(principal) => {
+            match crate::metering::reserve_or_reject(
+                Arc::clone(&fleet.entitlements),
+                principal,
+                bound,
+            )
+            .await
+            {
+                Ok(guard) => Some(guard),
+                Err(env) => return crate::error::envelope_response(env),
+            }
+        }
+        None => None,
+    };
+
+    let mut forward_headers = strip_hop_headers(&headers);
+    match axum::http::HeaderValue::from_str(&route.resolved_model_id) {
+        Ok(v) => {
+            forward_headers.insert(DECISION_MODEL_HEADER, v);
+        }
+        Err(_) => {
+            // A model id that is not a valid header value cannot have
+            // come from the catalogue or a neuron's model list.
+            return error_response(500, "api_error", "invalid_model_id", "internal error");
+        }
+    }
+
+    let url = format!("{}/v1/systemone", route.endpoint);
+    let start = Instant::now();
+    let upstream = fleet
+        .http_client
+        .post(&url)
+        .headers(forward_headers)
+        .header("content-type", "application/json")
+        .body(body)
+        .send()
+        .await;
+    let upstream = match upstream {
+        Ok(r) => r,
+        Err(e) => {
+            metrics::counter!("cortex_request_errors_total", &labels).increment(1);
+            tracing::warn!(node = %route.node_name, url = %url, error = %e, "systemone proxy failed");
+            return error_response(
+                502,
+                "api_error",
+                "upstream_unreachable",
+                "failed to reach the serving node",
+            );
+        }
+    };
+    let status = upstream.status();
+    let passthrough: Vec<(axum::http::HeaderName, axum::http::HeaderValue)> = upstream
+        .headers()
+        .iter()
+        .filter(|(n, _)| {
+            matches!(
+                n.as_str(),
+                "server-timing" | "x-inference-time-ms" | "retry-after"
+            )
+        })
+        .map(|(n, v)| (n.clone(), v.clone()))
+        .collect();
+    let resp_bytes = match upstream.bytes().await {
+        Ok(b) => b,
+        Err(e) => {
+            metrics::counter!("cortex_request_errors_total", &labels).increment(1);
+            tracing::warn!(node = %route.node_name, error = %e, "systemone response read failed");
+            return error_response(
+                502,
+                "api_error",
+                "upstream_read_failed",
+                "failed to read the serving node's response",
+            );
+        }
+    };
+    metrics::histogram!("cortex_request_duration_seconds", &labels)
+        .record(start.elapsed().as_secs_f64());
+
+    if status.is_success() {
+        let input_tokens = serde_json::from_slice::<Value>(&resp_bytes)
+            .ok()
+            .and_then(|v| v.get("usage")?.get("input_tokens")?.as_u64());
+        if let Some(n) = input_tokens {
+            metrics::counter!("cortex_decision_input_tokens_total", &labels).increment(n);
+        }
+        if let (Some(principal), Some(guard)) = (principal, guard) {
+            // No usage block means the node answered outside the
+            // contract; settle at the reserved bound rather than free.
+            let spent = input_tokens.unwrap_or_else(|| {
+                tracing::warn!(node = %route.node_name, "systemone response carried no usage");
+                bound
+            });
+            crate::metering::record_spend(&principal, spent, 0);
+            fleet
+                .served_usage
+                .add(&principal.account_id, &principal.key_id, spent);
+            guard.settle(spent);
+        }
+    } else {
+        metrics::counter!("cortex_request_errors_total", &labels).increment(1);
+        // guard drops here -> reservation released, no spend recorded.
+    }
+
+    let mut response = axum::response::Response::builder()
+        .status(status)
+        .header(axum::http::header::CONTENT_TYPE, "application/json");
+    for (name, value) in passthrough {
+        response = response.header(name, value);
+    }
+    response
+        .body(axum::body::Body::from(resp_bytes))
+        .unwrap_or_else(|_| {
+            error_response(500, "api_error", "response_build_failed", "internal error")
+        })
+}
+
 /// Forward client headers minus hop-by-hop and host headers, keeping
 /// the cortex-stamped principal headers intact for neuron's fair-share
 /// admission (#54).

crates/cortex-gateway/src/metering.rs

@@ -239,6 +239,35 @@ fn budget_error_to_envelope(err: BudgetError) -> OpenAiError {
     }
 }
 
+/// Sequence cap assumed for a decision request that does not set
+/// `max_len`: the largest default any shipped Laya checkpoint uses.
+const DECISION_DEFAULT_MAX_LEN: u64 = 1024;
+
+/// The ceiling a decision request may raise `max_len` to (Laya's own
+/// server caps the budget at the same value).
+const DECISION_MAX_LEN_CAP: u64 = 8192;
+
+/// Upper-bound input tokens for a `/v1/systemone` request (#338/#340).
+///
+/// A decision model encodes one sequence per question, each at most
+/// `max_len` tokens, and bills input tokens only — so questions ×
+/// `max_len` is a true ceiling, not an estimate. The node reports the
+/// exact count in `usage.input_tokens` and settle corrects to it.
+pub fn decision_reservation_tokens(body: &[u8]) -> u64 {
+    let v: serde_json::Value = serde_json::from_slice(body).unwrap_or_default();
+    let questions = match v.get("questions") {
+        Some(serde_json::Value::Object(m)) => m.len() as u64,
+        Some(serde_json::Value::Array(a)) => a.len() as u64,
+        _ => 1,
+    };
+    let max_len = v
+        .get("max_len")
+        .and_then(serde_json::Value::as_u64)
+        .unwrap_or(DECISION_DEFAULT_MAX_LEN)
+        .min(DECISION_MAX_LEN_CAP);
+    questions.max(1).saturating_mul(max_len.max(1))
+}
+
 /// Upper-bound tokens to reserve for a request (#52): an over-estimate of
 /// the prompt plus the maximum output. `advertised_output` is the model's
 /// `limit.output` (#62), used when the request omits `max_(completion_)tokens`.
@@ -271,6 +300,33 @@ fn estimate_prompt_tokens(body: &[u8]) -> u64 {
 mod tests {
     use super::*;
 
+    #[test]
+    fn decision_reservation_is_questions_times_sequence_cap() {
+        let body = br#"{"state":"x","questions":{"a":{},"b":{},"c":{}}}"#;
+        assert_eq!(
+            decision_reservation_tokens(body),
+            3 * DECISION_DEFAULT_MAX_LEN
+        );
+        let body = br#"{"state":"x","questions":{"a":{},"b":{}},"max_len":256}"#;
+        assert_eq!(decision_reservation_tokens(body), 512);
+    }
+
+    #[test]
+    fn decision_reservation_caps_a_raised_max_len() {
+        let body = br#"{"questions":{"a":{}},"max_len":1000000}"#;
+        assert_eq!(decision_reservation_tokens(body), DECISION_MAX_LEN_CAP);
+    }
+
+    #[test]
+    fn decision_reservation_never_zero() {
+        assert_eq!(
+            decision_reservation_tokens(b"not json"),
+            DECISION_DEFAULT_MAX_LEN
+        );
+        let body = br#"{"questions":{},"max_len":0}"#;
+        assert_eq!(decision_reservation_tokens(body), 1);
+    }
+
     #[test]
     fn requested_max_output_prefers_max_completion_tokens() {
         let body = br#"{"model":"m","max_completion_tokens":256,"max_tokens":99}"#;

crates/cortex-gateway/tests/systemone.rs

@@ -0,0 +1,274 @@
+//! `/v1/systemone` decision proxy tests (#338).
+//!
+//! The mock neuron echoes what it received — the `model` field and the
+//! routed-model header — so each test can assert both halves of the
+//! contract: *where* cortex sent the request, and that it did not
+//! rewrite the body on the way.
+
+use axum::Router;
+use axum::http::HeaderMap;
+use axum::response::{IntoResponse, Json};
+use axum::routing::{get, post};
+use cortex_core::catalogue::{DECISION_DEFAULT_ALIAS, DECISION_MODEL_HEADER};
+use cortex_core::config::{
+    ApiKeyConfig, EntitlementsConfig, EvictionSettings, EvictionStrategy, GatewayConfig,
+    GatewaySettings, NeuronEndpoint,
+};
+use cortex_core::entitlements::CapWindow;
+use cortex_core::node::{ModelEntry, ModelStatus};
+use cortex_gateway::state::CortexState;
+use serde_json::{Value, json};
+use std::sync::Arc;
+use tokio::net::TcpListener;
+
+const DECISION_MODEL: &str = "convaiinnovations/laya";
+const TEXT_MODEL: &str = "Qwen/Qwen3-1.7B";
+
+/// Mock neuron: `/v1/systemone` answers with a Jev-shaped body that
+/// echoes the forwarded `model` and routed-model header, and reports
+/// `usage.input_tokens` = 41 per question.
+async fn spawn_decision_mock_neuron() -> String {
+    let listener = TcpListener::bind("127.0.0.1:0").await.unwrap();
+    let addr = listener.local_addr().unwrap();
+    let base_url = format!("http://{addr}");
+    let inference_url = base_url.clone();
+    let app = Router::new()
+        .route(
+            "/models/{*rest}",
+            get(move || {
+                let url = inference_url.clone();
+                async move { Json(json!({"url": url})) }
+            }),
+        )
+        .route(
+            "/v1/systemone",
+            post(|headers: HeaderMap, Json(body): Json<Value>| async move {
+                let routed = headers
+                    .get(DECISION_MODEL_HEADER)
+                    .and_then(|v| v.to_str().ok())
+                    .map(str::to_string);
+                let questions = body
+                    .get("questions")
+                    .and_then(Value::as_object)
+                    .map(|m| m.len() as u64)
+                    .unwrap_or(0);
+                (
+                    [("server-timing", "inference;dur=12.50")],
+                    Json(json!({
+                        "model": "laya-rl-agent",
+                        "answers": {},
+                        "usage": {"input_tokens": 41 * questions, "output_tokens": 0},
+                        "echo_model": body.get("model").cloned().unwrap_or(Value::Null),
+                        "echo_routed": routed,
+                    })),
+                )
+                    .into_response()
+            }),
+        );
+    tokio::spawn(async move {
+        axum::serve(listener, app).await.unwrap();
+    });
+    base_url
+}
+
+fn entry(id: &str, capabilities: &[&str]) -> ModelEntry {
+    ModelEntry {
+        id: id.into(),
+        status: ModelStatus::Loaded,
+        last_accessed: None,
+        vram_estimate_mb: Some(1000),
+        capabilities: capabilities.iter().map(|c| c.to_string()).collect(),
+        tool_call: false,
+        reasoning: false,
+        limit: None,
+        servable: None,
+        reasoning_budget: Vec::new(),
+    }
+}
+
+/// A gateway fronting the mock with one decision and one text model
+/// loaded. `default_alias` controls whether `helexa/one` is configured;
+/// `hard_cap` switches on keyed auth with that balance.
+async fn spawn_gateway(neuron_url: &str, default_alias: bool, hard_cap: Option<u64>) -> String {
+    let entitlements = match hard_cap {
+        Some(cap) => EntitlementsConfig {
+            require_auth: true,
+            keys: vec![ApiKeyConfig {
+                key: "sk-dec".into(),
+                account_id: "acct-dec".into(),
+                key_id: Some("key-dec".into()),
+                hard_cap: Some(cap),
+                window: CapWindow::Balance,
+            }],
+        },
+        None => EntitlementsConfig::default(),
+    };
+    let config = GatewayConfig {
+        gateway: GatewaySettings {
+            listen: "127.0.0.1:0".into(),
+            metrics_listen: "127.0.0.1:0".into(),
+        },
+        eviction: EvictionSettings {
+            strategy: EvictionStrategy::Lru,
+            defrag_after_cycles: 0,
+        },
+        neurons: vec![NeuronEndpoint {
+            name: "mock-node".into(),
+            endpoint: neuron_url.to_string(),
+        }],
+        models_config: "/dev/null".into(),
+        entitlements,
+        upstream: Default::default(),
+    };
+    let mut state = CortexState::from_config(&config);
+    if default_alias {
+        state
+            .catalogue
+            .aliases
+            .insert(DECISION_DEFAULT_ALIAS.into(), DECISION_MODEL.into());
+    }
+    let fleet = Arc::new(state);
+    {
+        let mut nodes = fleet.nodes.write().await;
+        let node = nodes.get_mut("mock-node").unwrap();
+        node.healthy = true;
+        node.models
+            .insert(DECISION_MODEL.into(), entry(DECISION_MODEL, &["decision"]));
+        node.models
+            .insert(TEXT_MODEL.into(), entry(TEXT_MODEL, &[]));
+    }
+    let app = cortex_gateway::build_app(Arc::clone(&fleet));
+    let listener = TcpListener::bind("127.0.0.1:0").await.unwrap();
+    let addr = listener.local_addr().unwrap();
+    tokio::spawn(async move {
+        axum::serve(listener, app).await.unwrap();
+    });
+    format!("http://{addr}")
+}
+
+async fn send(gateway: &str, body: Value) -> (u16, HeaderMap, Value) {
+    let resp = reqwest::Client::new()
+        .post(format!("{gateway}/v1/systemone"))
+        .json(&body)
+        .send()
+        .await
+        .expect("request should reach the gateway");
+    let status = resp.status().as_u16();
+    let mut headers = HeaderMap::new();
+    for (k, v) in resp.headers() {
+        headers.insert(k.clone(), v.clone());
+    }
+    (status, headers, resp.json().await.unwrap())
+}
+
+fn request(model: Option<&str>) -> Value {
+    let mut body = json!({
+        "state": "I was charged twice this month",
+        "questions": {"refund": {"type": "noul", "instructions": "Refund request?"}},
+    });
+    if let Some(m) = model {
+        body["model"] = json!(m);
+    }
+    body
+}
+
+#[tokio::test]
+async fn test_explicit_decision_model_routes_and_body_is_untouched() {
+    let gateway = spawn_gateway(&spawn_decision_mock_neuron().await, true, None).await;
+    let (status, headers, body) = send(&gateway, request(Some(DECISION_MODEL))).await;
+    assert_eq!(status, 200, "{body}");
+    assert_eq!(body["echo_routed"], DECISION_MODEL);
+    assert_eq!(body["echo_model"], DECISION_MODEL);
+    assert_eq!(body["usage"]["input_tokens"], 41);
+    // Upstream timing headers survive the hop.
+    assert_eq!(
+        headers.get("server-timing").and_then(|v| v.to_str().ok()),
+        Some("inference;dur=12.50")
+    );
+}
+
+#[tokio::test]
+async fn test_foreign_model_name_falls_back_to_default_alias() {
+    // A Jev client names its own service. The request must still be
+    // served — routed via helexa/one — and the client's value must reach
+    // the node unchanged (it may be a checkpoint pin the node honours).
+    let gateway = spawn_gateway(&spawn_decision_mock_neuron().await, true, None).await;
+    let (status, _, body) = send(&gateway, request(Some("jev-1"))).await;
+    assert_eq!(status, 200, "{body}");
+    assert_eq!(body["echo_routed"], DECISION_MODEL);
+    assert_eq!(body["echo_model"], "jev-1");
+}
+
+#[tokio::test]
+async fn test_missing_model_falls_back_to_default_alias() {
+    let gateway = spawn_gateway(&spawn_decision_mock_neuron().await, true, None).await;
+    let (status, _, body) = send(&gateway, request(None)).await;
+    assert_eq!(status, 200, "{body}");
+    assert_eq!(body["echo_routed"], DECISION_MODEL);
+    assert_eq!(body["echo_model"], Value::Null);
+}
+
+#[tokio::test]
+async fn test_default_alias_itself_resolves() {
+    let gateway = spawn_gateway(&spawn_decision_mock_neuron().await, true, None).await;
+    let (status, _, body) = send(&gateway, request(Some(DECISION_DEFAULT_ALIAS))).await;
+    assert_eq!(status, 200, "{body}");
+    assert_eq!(body["echo_routed"], DECISION_MODEL);
+}
+
+#[tokio::test]
+async fn test_known_non_decision_model_is_not_substituted() {
+    // A model the fleet does serve routes to itself; the node owns the
+    // wrong_modality answer. Silently swapping in the decision model
+    // would hide the client's mistake.
+    let gateway = spawn_gateway(&spawn_decision_mock_neuron().await, true, None).await;
+    let (status, _, body) = send(&gateway, request(Some(TEXT_MODEL))).await;
+    assert_eq!(status, 200, "{body}");
+    assert_eq!(body["echo_routed"], TEXT_MODEL);
+}
+
+#[tokio::test]
+async fn test_unknown_model_without_default_alias_is_404() {
+    let gateway = spawn_gateway(&spawn_decision_mock_neuron().await, false, None).await;
+    let (status, _, body) = send(&gateway, request(Some("jev-1"))).await;
+    assert_eq!(status, 404, "{body}");
+    assert_eq!(body["error"]["code"], "model_not_found");
+}
+
+#[tokio::test]
+async fn test_budget_below_reservation_rejected_before_dispatch() {
+    // One question reserves one full sequence (1024 tokens by default);
+    // a 100-token balance cannot cover it.
+    let gateway = spawn_gateway(&spawn_decision_mock_neuron().await, true, Some(100)).await;
+    let resp = reqwest::Client::new()
+        .post(format!("{gateway}/v1/systemone"))
+        .bearer_auth("sk-dec")
+        .json(&request(Some(DECISION_MODEL)))
+        .send()
+        .await
+        .unwrap();
+    assert_eq!(resp.status(), 429);
+    let body: Value = resp.json().await.unwrap();
+    assert_eq!(body["error"]["code"], "insufficient_quota");
+}
+
+#[tokio::test]
+async fn test_spend_settles_to_reported_input_tokens() {
+    // Balance for exactly two one-question requests at the reserved
+    // bound (2 × 1024). Settling to the reported 41 tokens leaves room
+    // for many more; settling at the bound would exhaust it on the third.
+    let gateway = spawn_gateway(&spawn_decision_mock_neuron().await, true, Some(2048)).await;
+    let client = reqwest::Client::new();
+    for i in 0..5 {
+        let resp = client
+            .post(format!("{gateway}/v1/systemone"))
+            .bearer_auth("sk-dec")
+            .json(&request(Some(DECISION_MODEL)))
+            .send()
+            .await
+            .unwrap();
+        assert_eq!(resp.status(), 200, "request {i} should be affordable");
+        // Settlement is spawned; let it land before the next reservation.
+        tokio::time::sleep(std::time::Duration::from_millis(50)).await;
+    }
+}

crates/helexa-bench/src/scenario.rs

@@ -153,17 +153,29 @@ pub trait Scenario: Send + Sync {
     fn prompt_size(&self) -> u32;
 
     /// Whether this scenario should run against the given model. The
-    /// default runs against everything except image-generation models
-    /// (a chat request at an image model is a 422 `wrong_modality`);
-    /// modality-specific scenarios override this.
+    /// default runs against chat models only: an image-generation model
+    /// or a decision model (#333) answers a chat request with a 422
+    /// `wrong_modality`. Modality-specific scenarios override this.
     fn applies_to(&self, model: &ModelInfo) -> bool {
-        !model.capabilities.iter().any(|c| c == "image")
+        serves_chat(model)
     }
 
     /// Issue one shaped request and measure it.
     async fn run(&self, ctx: &RunCtx) -> Result<ScenarioMetrics>;
 }
 
+/// Capabilities that mark a model as serving an endpoint other than
+/// chat completions.
+const NON_CHAT_MODALITIES: [&str; 2] = ["image", "decision"];
+
+/// Whether a model answers `/v1/chat/completions`.
+pub fn serves_chat(model: &ModelInfo) -> bool {
+    !model
+        .capabilities
+        .iter()
+        .any(|c| NON_CHAT_MODALITIES.contains(&c.as_str()))
+}
+
 /// Fixed-seed image-latency scenario (#203): one 9-step generation at a
 /// configured square size against `/v1/images/generations`, recording
 /// the server-side phase timing (`usage.helexa_timing`) and the metered
@@ -883,6 +895,34 @@ async fn stream_and_measure_inner(
 mod tests {
     use super::*;
 
+    fn model(capabilities: &[&str]) -> ModelInfo {
+        serde_json::from_value(serde_json::json!({
+            "id": "m",
+            "harness": "candle",
+            "status": "loaded",
+            "devices": [0],
+            "vram_used_mb": null,
+            "capabilities": capabilities,
+        }))
+        .expect("model info")
+    }
+
+    /// Chat scenarios skip every model that answers chat with
+    /// `wrong_modality`: image generators and decision models alike.
+    #[test]
+    fn chat_scenarios_skip_image_and_decision_models() {
+        assert!(serves_chat(&model(&[])));
+        assert!(serves_chat(&model(&["text", "vision"])));
+        assert!(!serves_chat(&model(&["image"])));
+        assert!(!serves_chat(&model(&["decision"])));
+        let image = ImageLatencyScenario {
+            id: "image:512".into(),
+            side_px: 512,
+        };
+        assert!(image.applies_to(&model(&["image"])));
+        assert!(!image.applies_to(&model(&["decision"])));
+    }
+
     /// A burst must report server-side phase timing, or `queue_wait_ms`
     /// (TTFT minus server prefill, #85) cannot be computed for the very
     /// cells it was designed for, and the prefill-tok/s panel stays

crates/helexa-router/src/dispatch.rs

@@ -149,25 +149,81 @@ pub async fn dispatch(
 
     let candidates = match select_cortexes(state, &model).await {
         Selection::Candidates(c) => c,
-        Selection::UnknownModel => {
-            return envelope_response(
-                OpenAiError::new(
-                    404,
-                    "invalid_request_error",
-                    "model_not_found",
-                    format!("no operator serves model '{model}'"),
-                )
-                .with_param("model"),
-            );
-        }
-        Selection::NoReachableCapacity => {
-            return envelope_response(OpenAiError::service_unavailable(
-                format!("model '{model}' is temporarily unavailable on all operators"),
-                Some(RETRY_AFTER_SECS),
-            ));
-        }
+        other => return selection_error(other, &model),
     };
+    forward(state, path, headers, body, &model, candidates).await
+}
 
+/// Proxy a `/v1/systemone` decision request (#338).
+///
+/// Same as [`dispatch`], except that a missing `model`, or one no
+/// operator serves, selects capacity for
+/// [`DECISION_DEFAULT_ALIAS`](cortex_core::catalogue::DECISION_DEFAULT_ALIAS)
+/// instead of 404ing: decision clients written against the Jev API name
+/// their own service in `model`. The body is forwarded with the client's
+/// `model` intact — cortex applies the same fallback and the serving node
+/// may read it as a checkpoint pin.
+pub async fn dispatch_decision(
+    state: &RouterState,
+    path: &str,
+    headers: HeaderMap,
+    body: Bytes,
+) -> Response {
+    use cortex_core::catalogue::DECISION_DEFAULT_ALIAS;
+
+    let requested = extract_model(&body);
+    let (mut model, body) = match requested.as_deref() {
+        Some(m) => match state.aliases.get(m) {
+            Some(real) => {
+                tracing::debug!(alias = m, model = %real, "resolving tier alias");
+                (real.clone(), rewrite_model_in_body(&body, real))
+            }
+            None => (m.to_string(), body),
+        },
+        None => (DECISION_DEFAULT_ALIAS.to_string(), body),
+    };
+    let mut selection = select_cortexes(state, &model).await;
+    if selection == Selection::UnknownModel
+        && let Some(default) = state.aliases.get(DECISION_DEFAULT_ALIAS)
+    {
+        tracing::debug!(requested = %model, default = %default, "unknown decision model; using the default");
+        model = default.clone();
+        selection = select_cortexes(state, &model).await;
+    }
+    match selection {
+        Selection::Candidates(c) => forward(state, path, headers, body, &model, c).await,
+        other => selection_error(other, &model),
+    }
+}
+
+/// The #63 envelope for a selection that produced no candidates.
+fn selection_error(selection: Selection, model: &str) -> Response {
+    match selection {
+        Selection::Candidates(_) | Selection::UnknownModel => envelope_response(
+            OpenAiError::new(
+                404,
+                "invalid_request_error",
+                "model_not_found",
+                format!("no operator serves model '{model}'"),
+            )
+            .with_param("model"),
+        ),
+        Selection::NoReachableCapacity => envelope_response(OpenAiError::service_unavailable(
+            format!("model '{model}' is temporarily unavailable on all operators"),
+            Some(RETRY_AFTER_SECS),
+        )),
+    }
+}
+
+/// Try `candidates` in order, failing over only on transport errors.
+async fn forward(
+    state: &RouterState,
+    path: &str,
+    headers: HeaderMap,
+    body: Bytes,
+    model: &str,
+    candidates: Vec<CortexEndpoint>,
+) -> Response {
     // Try candidates in order, failing over only on transport errors. A
     // genuine HTTP response (any status — including cortex's #63 429/400)
     // is returned verbatim and never retried away.

crates/helexa-router/src/handlers.rs

@@ -14,6 +14,7 @@ pub fn api_routes() -> Router<Arc<RouterState>> {
         .route("/v1/chat/completions", post(chat_completions))
         .route("/v1/completions", post(completions))
         .route("/v1/images/generations", post(images_generations))
+        .route("/v1/systemone", post(systemone))
         .route("/v1/responses", post(responses))
         .route("/v1/messages", post(messages))
         .route("/v1/models", get(list_models))
@@ -63,6 +64,17 @@ async fn images_generations(
     dispatch::dispatch(&state, "/v1/images/generations", headers, body).await
 }
 
+/// `POST /v1/systemone` (#338): typed decisions (TypeSafe Jev wire
+/// protocol) from a decision model. Unlike the other paths an unknown or
+/// missing `model` is not a 404 — see [`dispatch::dispatch_decision`].
+async fn systemone(
+    State(state): State<Arc<RouterState>>,
+    headers: HeaderMap,
+    body: Bytes,
+) -> Response {
+    dispatch::dispatch_decision(&state, "/v1/systemone", headers, body).await
+}
+
 async fn messages(
     State(state): State<Arc<RouterState>>,
     headers: HeaderMap,

crates/helexa-router/tests/dispatch.rs

@@ -14,7 +14,7 @@ use axum::routing::post;
 use axum::{Json, Router};
 use cortex_core::node::CortexModelEntry;
 use helexa_router::config::{CortexEndpoint, RouterConfig};
-use helexa_router::dispatch::{Selection, dispatch, select_cortexes};
+use helexa_router::dispatch::{Selection, dispatch, dispatch_decision, select_cortexes};
 use helexa_router::state::{CortexTopology, RouterState};
 use serde_json::{Value, json};
 use std::collections::HashMap;
@@ -87,6 +87,7 @@ async fn mock_handler(State(m): State<MockCortex>, headers: HeaderMap, body: Byt
 async fn spawn_cortex(mock: MockCortex) -> String {
     let app = Router::new()
         .route("/v1/chat/completions", post(mock_handler))
+        .route("/v1/systemone", post(mock_handler))
         .with_state(mock);
     let listener = TcpListener::bind("127.0.0.1:0").await.unwrap();
     let addr = listener.local_addr().unwrap();
@@ -349,3 +350,75 @@ async fn ranking_prefers_loaded_then_region_then_headroom() {
     let names: Vec<&str> = order.iter().map(|e| e.name.as_str()).collect();
     assert_eq!(names, vec!["warm-eu", "warm-us", "cold-eu"]);
 }
+
+// ── /v1/systemone default-model fallback (#338) ──────────────────────
+
+/// Router state with c1 serving MODEL, and `helexa/one` pointing at it
+/// when `with_default` is set.
+async fn decision_state(with_default: bool) -> RouterState {
+    let url = spawn_cortex(ok_cortex("c1")).await;
+    let mut state = state_with(vec![ep("c1", &url, None)], None);
+    if with_default {
+        state.aliases.insert(
+            cortex_core::catalogue::DECISION_DEFAULT_ALIAS.to_string(),
+            MODEL.to_string(),
+        );
+    }
+    set_topology(&state, "c1", true, true, true, 2).await;
+    state
+}
+
+#[tokio::test]
+async fn decision_foreign_model_selects_default_and_keeps_body_model() {
+    // A Jev client names its own service. Capacity is chosen for the
+    // default decision model, and the client's `model` is forwarded
+    // untouched for cortex (and the node) to interpret.
+    let state = decision_state(true).await;
+    let body = Bytes::from_static(b"{\"model\":\"jev-1\",\"questions\":{}}");
+    let resp = dispatch_decision(&state, "/v1/systemone", HeaderMap::new(), body).await;
+    let (status, body) = body_json(resp).await;
+    assert_eq!(status, StatusCode::OK);
+    assert_eq!(body["served_by"], "c1");
+    assert_eq!(body["model_seen"], "jev-1");
+}
+
+#[tokio::test]
+async fn decision_missing_model_selects_default() {
+    let state = decision_state(true).await;
+    let body = Bytes::from_static(b"{\"questions\":{}}");
+    let resp = dispatch_decision(&state, "/v1/systemone", HeaderMap::new(), body).await;
+    let (status, body) = body_json(resp).await;
+    assert_eq!(status, StatusCode::OK);
+    assert_eq!(body["served_by"], "c1");
+    assert_eq!(body["model_seen"], "");
+}
+
+#[tokio::test]
+async fn decision_default_alias_is_rewritten_like_any_alias() {
+    let state = decision_state(true).await;
+    let body = Bytes::from_static(b"{\"model\":\"helexa/one\",\"questions\":{}}");
+    let resp = dispatch_decision(&state, "/v1/systemone", HeaderMap::new(), body).await;
+    let (status, body) = body_json(resp).await;
+    assert_eq!(status, StatusCode::OK);
+    assert_eq!(body["model_seen"], MODEL);
+}
+
+#[tokio::test]
+async fn decision_unknown_model_without_default_is_404() {
+    let state = decision_state(false).await;
+    let body = Bytes::from_static(b"{\"model\":\"jev-1\",\"questions\":{}}");
+    let resp = dispatch_decision(&state, "/v1/systemone", HeaderMap::new(), body).await;
+    let (status, body) = body_json(resp).await;
+    assert_eq!(status, StatusCode::NOT_FOUND);
+    assert_eq!(body["error"]["code"], "model_not_found");
+}
+
+#[tokio::test]
+async fn chat_unknown_model_does_not_take_the_decision_fallback() {
+    // The fallback is scoped to the decision endpoint.
+    let state = decision_state(true).await;
+    let body = Bytes::from_static(b"{\"model\":\"jev-1\",\"stream\":false}");
+    let resp = dispatch(&state, "/v1/chat/completions", HeaderMap::new(), body).await;
+    let (status, _) = body_json(resp).await;
+    assert_eq!(status, StatusCode::NOT_FOUND);
+}

crates/neuron/Cargo.toml

@@ -12,6 +12,14 @@ path = "src/lib.rs"
 name = "neuron"
 path = "src/main.rs"
 
+# Measurement harness for the MTP speculative-decoding stages (#96).
+# Not shipped in the RPM: it loads a checkpoint directly and reports
+# numbers, and exists so the go/no-go gate is a measurement rather than
+# an argument.
+[[bin]]
+name = "neuron-mtp-probe"
+path = "src/bin/mtp_probe.rs"
+
 [features]
 default = []
 # Enables CUDA acceleration in candle and the cudarc/nccl bindings the
@@ -47,6 +55,8 @@ cuda-integration = ["cuda"]
 cortex-core.workspace = true
 tokio.workspace = true
 axum.workspace = true
+metrics.workspace = true
+metrics-exporter-prometheus.workspace = true
 serde.workspace = true
 serde_json.workspace = true
 reqwest.workspace = true
@@ -66,6 +76,12 @@ toml.workspace = true
 # transitively via candle-core.
 rayon = "1"
 
+# Unicode tables for the decision-model language detector (#339), which
+# must classify characters exactly as Python's `str.isalpha` / `re` and
+# `unicodedata.combining` do. Both already in the tree via tokenizers.
+unicode-normalization = "0.1"
+unicode-properties = "0.1"
+
 # candle for in-process inference. CUDA support is gated behind the
 # crate's `cuda` feature (default off) so the workspace builds on
 # non-CUDA hosts and CI runners.

crates/neuron/src/api.rs

@@ -10,6 +10,7 @@ use crate::wire::{openai_chat, openai_responses};
 use axum::Router;
 use axum::extract::{Path, State};
 use axum::http::StatusCode;
+use axum::http::header;
 use axum::response::sse::{Event, KeepAlive, Sse};
 use axum::response::{IntoResponse, Json};
 use axum::routing::{get, post};
@@ -43,6 +44,7 @@ pub struct NeuronState {
 /// Build the neuron API router.
 pub fn neuron_routes() -> Router<Arc<NeuronState>> {
     Router::new()
+        .route("/metrics", get(metrics_handler))
         .route("/version", get(version_handler))
         .route("/discovery", get(discovery_handler))
         .route("/health", get(health_handler))
@@ -53,6 +55,7 @@ pub fn neuron_routes() -> Router<Arc<NeuronState>> {
         .route("/models/{model_id}/capabilities", get(model_capabilities))
         .route("/v1/chat/completions", post(chat_completions))
         .route("/v1/images/generations", post(images_generations))
+        .route("/v1/systemone", post(systemone))
         .route("/v1/responses", post(responses))
 }
 
@@ -64,6 +67,29 @@ async fn version_handler() -> Json<cortex_core::build_info::BuildInfo> {
     Json(crate::version::build_info())
 }
 
+/// `GET /metrics` — Prometheus text format.
+///
+/// Served from the API port rather than a second listener, so a scrape
+/// target needs no extra firewall rule. A missing recorder is a 503
+/// rather than an empty 200: an empty scrape looks exactly like a
+/// healthy process with nothing to say, and would leave the fleet
+/// silently unmonitored.
+async fn metrics_handler() -> impl IntoResponse {
+    match crate::metrics::render() {
+        Some(body) => (
+            StatusCode::OK,
+            [(header::CONTENT_TYPE, "text/plain; version=0.0.4")],
+            body,
+        )
+            .into_response(),
+        None => (
+            StatusCode::SERVICE_UNAVAILABLE,
+            "metrics recorder not installed",
+        )
+            .into_response(),
+    }
+}
+
 async fn discovery_handler(State(state): State<Arc<NeuronState>>) -> Json<DiscoveryResponse> {
     Json(state.discovery.clone())
 }
@@ -164,6 +190,7 @@ fn preflight_kind(err: &PreflightError) -> &'static str {
         PreflightError::EmptyRepo { .. } => "empty_repo",
         PreflightError::TpRequiresSafetensors { .. } => "tp_requires_safetensors",
         PreflightError::QuantNotFound { .. } => "quant_not_found",
+        PreflightError::Unsupported { .. } => "unsupported_load_option",
     }
 }
 
@@ -643,6 +670,83 @@ async fn images_generations(
     }
 }
 
+/// `POST /v1/systemone` (#338): typed decisions over the TypeSafe Jev
+/// wire protocol, from a loaded decision model (Laya). Parsing and the
+/// transport-level limits happen here; routing, validation, admission,
+/// scoring and shaping in `CandleHarness::systemone`.
+///
+/// Errors carry the #63 envelope *and* the reference server's top-level
+/// `detail` string, so OpenAI-style and Jev/laya clients both read them.
+async fn systemone(
+    State(state): State<Arc<NeuronState>>,
+    headers: axum::http::HeaderMap,
+    body: axum::body::Bytes,
+) -> axum::response::Response {
+    use crate::harness::candle::SystemOneError;
+    use cortex_core::decisions::{SystemOneRequest, limits};
+
+    let Some(candle) = state.candle.as_ref().map(Arc::clone) else {
+        return decision_error_response(503, "candle harness not enabled on this neuron");
+    };
+    if body.len() > limits::MAX_BODY_BYTES {
+        return decision_error_response(413, "request body too large");
+    }
+    let request = match SystemOneRequest::parse(&body, limits::DEFAULT_MAX_TOKEN_BUDGET) {
+        Ok(r) => r,
+        Err(e) => return decision_error_response(e.status, &e.detail),
+    };
+    let served = headers
+        .get(cortex_core::catalogue::DECISION_MODEL_HEADER)
+        .and_then(|v| v.to_str().ok())
+        .map(str::to_string);
+    let principal = principal_key(&headers);
+
+    let started = std::time::Instant::now();
+    match candle
+        .systemone(served.as_deref(), request, principal.as_deref())
+        .await
+    {
+        Ok(response) => {
+            let ms = started.elapsed().as_secs_f64() * 1e3;
+            let mut out = Json(response).into_response();
+            for (name, value) in [
+                ("server-timing", format!("inference;dur={ms:.2}")),
+                ("x-inference-time-ms", format!("{ms:.2}")),
+            ] {
+                if let Ok(v) = axum::http::HeaderValue::from_str(&value) {
+                    out.headers_mut().insert(name, v);
+                }
+            }
+            out
+        }
+        Err(SystemOneError::Request(e)) => decision_error_response(e.status, &e.detail),
+        Err(SystemOneError::Inference(e)) => {
+            tracing::warn!(error = %e, "systemone request failed");
+            inference_error_response(e)
+        }
+    }
+}
+
+/// A rejected decision request: the #63 envelope, plus the `detail` field
+/// the reference (FastAPI) server answers with.
+fn decision_error_response(status: u16, detail: &str) -> axum::response::Response {
+    use cortex_core::error_envelope::OpenAiError;
+    let (kind, code) = match status {
+        404 => ("invalid_request_error", "model_not_found"),
+        413 => ("invalid_request_error", "request_too_large"),
+        422 => ("invalid_request_error", "invalid_decision_request"),
+        400..=499 => ("invalid_request_error", "invalid_request"),
+        _ => ("api_error", "service_unavailable"),
+    };
+    let env = OpenAiError::new(status, kind, code, detail.to_string());
+    let mut body = env.body();
+    if let Some(obj) = body.as_object_mut() {
+        obj.insert("detail".into(), json!(detail));
+    }
+    let status = StatusCode::from_u16(status).unwrap_or(StatusCode::BAD_REQUEST);
+    (status, Json(body)).into_response()
+}
+
 fn finish_reason_from_str(s: &str) -> crate::wire::FinishReason {
     use crate::wire::FinishReason;
     match s {

crates/neuron/src/bin/mtp_probe.rs

@@ -0,0 +1,730 @@
+//! MTP speculative-decoding measurement harness (#96, stages S3/S4).
+//!
+//! Two numbers decide whether speculative decoding is worth building
+//! into neuron, and neither can be taken from anyone else's benchmark:
+//!
+//! - **`c`** — what a draft step costs relative to a target decode
+//!   step. Published figures are for a different drafter on different
+//!   hardware. Ours reads the target's own `lm_head` (~1.03 GB at q6k
+//!   on the 27B) on top of the head's weights, so the vocabulary, not
+//!   the head, may dominate.
+//! - **acceptance** — how many drafted tokens the target actually
+//!   agrees with, *on our workloads*. Every published figure is
+//!   short-form (GSM8K, MATH-500, HumanEval, ≤4k output); helexa serves
+//!   15–50k-token agentic contexts full of replayed reasoning, and
+//!   acceptance is workload-dependent. A poor rate makes speculation a
+//!   net loss.
+//!
+//! This binary answers both by running a real checkpoint. It commits
+//! nothing, changes no serving behaviour, and is not shipped in the
+//! RPM.
+//!
+//! ## What it does
+//!
+//! Prefills the target over a prompt, then generates greedily. At each
+//! step it snapshots the draft head's KV, drafts `K` tokens by chaining
+//! the head, restores the snapshot, and lets the target continue
+//! normally. When the target has produced the next `K` true tokens, the
+//! draft is scored against them: the accepted length is the number of
+//! leading matches.
+//!
+//! The head is *not* the thing generating — the target's own output is
+//! never influenced by the draft, which is what makes this an observer
+//! rather than an implementation.
+//!
+//! ## Two logits, two argmaxes
+//!
+//! Acceptance is reported against the target's raw argmax **and**
+//! against its argmax after the repeat penalty neuron applies when
+//! serving. The two differ, and the gap is exactly what a
+//! distribution-preserving (rejection-sampling) scheme would have to
+//! reconcile at `temperature > 0` — this fleet serves this family at
+//! 0.6. Measuring it here costs nothing and makes that decision an
+//! informed one.
+//!
+//! ## Usage
+//!
+//! ```sh
+//! neuron-mtp-probe --model /path/to/Qwen3.5-0.8B --device cuda:0 \
+//!     --draft-len 4 --steps 64 --prompt-file prompt.txt
+//! ```
+
+use anyhow::{Context, Result};
+use candle_core::{DType, Device, IndexOp, Tensor};
+use clap::Parser;
+use std::path::PathBuf;
+use std::time::Instant;
+
+use neuron::harness::arch::qwen3_5::mtp::MtpHead;
+use neuron::harness::arch::qwen3_5::rope::RotaryEmbedding;
+use neuron::harness::arch::qwen3_5::{Config, Qwen3_5ForCausalLM};
+
+#[derive(Parser, Debug)]
+#[command(about = "Measure MTP draft cost and acceptance rate (#96)")]
+struct Args {
+    /// Checkpoint directory: config.json, *.safetensors, tokenizer.json.
+    #[arg(long)]
+    model: PathBuf,
+    /// `cpu` or `cuda:N`.
+    #[arg(long, default_value = "cpu")]
+    device: String,
+    /// Draft tokens per round (K).
+    #[arg(long, default_value_t = 4)]
+    draft_len: usize,
+    /// Decode steps to measure over.
+    #[arg(long, default_value_t = 64)]
+    steps: usize,
+    /// Prompt file. Without one, a short built-in prompt is used — fine
+    /// for `c`, useless for acceptance, which is the whole point of
+    /// measuring on a replayed session instead.
+    #[arg(long)]
+    prompt_file: Option<PathBuf>,
+    /// Prefill chunk size. The mask for a one-shot prefill is L x L —
+    /// 2.5 GB at 25k tokens — so the probe chunks like the serving path
+    /// does.
+    #[arg(long, default_value_t = 512)]
+    prefill_chunk: usize,
+    /// Timed-but-discarded rounds before measurement starts.
+    #[arg(long, default_value_t = 3)]
+    warmup: usize,
+    /// Repeat penalty applied to the target's logits when scoring the
+    /// second argmax. Mirrors the serving default.
+    #[arg(long, default_value_t = 1.05)]
+    repeat_penalty: f32,
+    /// How many recent tokens the repeat penalty considers.
+    #[arg(long, default_value_t = 64)]
+    repeat_last_n: usize,
+    /// Tensor-parallel degree. 1 loads the model on one card; 2+ uses
+    /// the worker pool, which is the only way to reach a model too big
+    /// for a single GPU — the 27B is 54 GB in bf16 and the single-GPU
+    /// path is bf16-only.
+    #[arg(long, default_value_t = 1)]
+    tensor_parallel: u32,
+    /// Quantisation for the TP load (`q6k`, `q8_0`, …). Only the TP
+    /// loader applies it.
+    #[arg(long)]
+    quant: Option<String>,
+    /// The binary the worker ranks are spawned from. Defaults to
+    /// `neuron` beside this executable — the probe itself does not
+    /// implement `--worker`.
+    #[arg(long)]
+    worker_binary: Option<PathBuf>,
+    /// How many verify-shaped forwards to time after the measurement
+    /// loop. A single sample moved 57.3 -> 50.1 ms between otherwise
+    /// identical runs — ±7 ms on a 45 ms baseline, enough to move the
+    /// predicted speedup across the gate on its own.
+    #[arg(long, default_value_t = 9)]
+    verify_samples: usize,
+    /// Write the summary as JSON here as well as to stdout.
+    #[arg(long)]
+    json: Option<PathBuf>,
+}
+
+fn pick_device(spec: &str) -> Result<Device> {
+    if spec == "cpu" {
+        return Ok(Device::Cpu);
+    }
+    let idx: usize = spec
+        .strip_prefix("cuda:")
+        .ok_or_else(|| anyhow::anyhow!("--device must be 'cpu' or 'cuda:N', got '{spec}'"))?
+        .parse()
+        .context("parse cuda device index")?;
+    #[cfg(feature = "cuda")]
+    {
+        Ok(Device::new_cuda(idx)?)
+    }
+    #[cfg(not(feature = "cuda"))]
+    {
+        let _ = idx;
+        anyhow::bail!("this binary was built without the `cuda` feature")
+    }
+}
+
+fn argmax(row: &[f32]) -> u32 {
+    let mut best = 0usize;
+    for (i, v) in row.iter().enumerate() {
+        if v > &row[best] {
+            best = i;
+        }
+    }
+    best as u32
+}
+
+/// The target's argmax after the repeat penalty neuron applies when
+/// serving. Same shape of penalty as `sample_with_penalty`: recent
+/// tokens are divided (positive logits) or multiplied (negative ones).
+fn argmax_with_penalty(row: &[f32], recent: &[u32], penalty: f32) -> u32 {
+    if penalty == 1.0 {
+        return argmax(row);
+    }
+    let mut adjusted = row.to_vec();
+    for &t in recent {
+        let i = t as usize;
+        if i < adjusted.len() {
+            adjusted[i] = if adjusted[i] > 0.0 {
+                adjusted[i] / penalty
+            } else {
+                adjusted[i] * penalty
+            };
+        }
+    }
+    argmax(&adjusted)
+}
+
+/// Additive causal mask, `(1, 1, l, l + offset)`.
+fn causal_mask(l: usize, offset: usize, dtype: DType, dev: &Device) -> Result<Tensor> {
+    let total = l + offset;
+    let mut data = vec![0f32; l * total];
+    for i in 0..l {
+        for j in 0..total {
+            if j > i + offset {
+                data[i * total + j] = f32::NEG_INFINITY;
+            }
+        }
+    }
+    Ok(Tensor::from_vec(data, (1, 1, l, total), dev)?.to_dtype(dtype)?)
+}
+
+/// Walk the prompt through target and head, in chunks.
+///
+/// Two reasons this is chunked rather than one forward: the causal mask
+/// is `L x L` — 2.5 GB at 25k tokens, which OOMs a 12 GB card before a
+/// single weight is touched — and the serving path chunks too, so this
+/// keeps the measurement on the same shape of work.
+///
+/// The head trails the target by one token: at position `i` it takes
+/// the embedding of token `i+1` with the target's hidden state at `i`.
+/// Each chunk therefore reaches one token into the next. The final
+/// prompt position has no successor yet — the first generated token
+/// fills that slot.
+///
+/// Returns the target's logits and hidden state at the last position.
+#[allow(clippy::too_many_arguments)]
+fn prefill(
+    model: &mut Qwen3_5ForCausalLM,
+    head: &mut MtpHead,
+    rotary: &RotaryEmbedding,
+    ids: &[u32],
+    chunk: usize,
+    dtype: DType,
+    dev: &Device,
+    sync: &dyn Fn(&Device) -> Result<()>,
+) -> Result<(Tensor, Tensor)> {
+    let n = ids.len();
+    let mut last: Option<(Tensor, Tensor)> = None;
+    let mut at = 0usize;
+    while at < n {
+        let end = (at + chunk).min(n);
+        let block = Tensor::new(&ids[at..end], dev)?.unsqueeze(0)?;
+        let (logits, hidden) = model.forward_with_hidden(&block, at)?;
+        let l = end - at;
+
+        let head_end = end.min(n - 1);
+        if head_end > at {
+            let hl = head_end - at;
+            let shifted = Tensor::new(&ids[at + 1..head_end + 1], dev)?.unsqueeze(0)?;
+            let embeds = model.embed_tokens(&shifted)?;
+            let hid = hidden.i((.., ..hl, ..))?;
+            let mask = causal_mask(hl, at, dtype, dev)?;
+            let positions: Vec<usize> = (at..head_end).collect();
+            let (cos, sin) = rotary.cos_sin_at(&positions)?;
+            head.forward(&embeds, &hid, Some(&mask), &cos, &sin)?;
+        }
+
+        last = Some((logits, hidden.i((.., l - 1.., ..))?));
+        at = end;
+        sync(dev)?;
+    }
+    last.context("empty prompt")
+}
+
+#[tokio::main]
+async fn main() -> Result<()> {
+    let args = Args::parse();
+    if args.tensor_parallel > 1 {
+        return run_tp(args).await;
+    }
+    let dev = pick_device(&args.device)?;
+    let dtype = if matches!(dev, Device::Cpu) {
+        DType::F32
+    } else {
+        DType::BF16
+    };
+
+    // ── load ────────────────────────────────────────────────────────
+    let cfg_raw =
+        std::fs::read_to_string(args.model.join("config.json")).context("read config.json")?;
+    let config: Config = Config::from_config_json(&cfg_raw).context("parse config.json")?;
+    let text_cfg = config.text_config.clone();
+    anyhow::ensure!(
+        MtpHead::present_in(&text_cfg),
+        "this checkpoint declares no MTP head (mtp_num_hidden_layers = 0)"
+    );
+
+    let mut shards: Vec<PathBuf> = std::fs::read_dir(&args.model)
+        .context("read model dir")?
+        .filter_map(|e| e.ok().map(|e| e.path()))
+        .filter(|p| p.extension().is_some_and(|x| x == "safetensors"))
+        .collect();
+    shards.sort();
+    anyhow::ensure!(!shards.is_empty(), "no .safetensors in {:?}", args.model);
+    eprintln!(
+        "loading {} shard(s) as {dtype:?} on {:?}…",
+        shards.len(),
+        dev
+    );
+
+    // SAFETY: mmaps checkpoint files the operator pointed us at.
+    let vb =
+        unsafe { candle_nn::var_builder::ShardedSafeTensors::var_builder(&shards, dtype, &dev)? };
+    let mut model = Qwen3_5ForCausalLM::new(config, vb.clone()).context("load target model")?;
+    let rotary = std::sync::Arc::new(RotaryEmbedding::new(dtype, &text_cfg, &dev)?);
+    let mut head = MtpHead::load(&text_cfg, rotary.clone(), &vb).context("load MTP head")?;
+
+    let tok = tokenizers::Tokenizer::from_file(args.model.join("tokenizer.json"))
+        .map_err(|e| anyhow::anyhow!("load tokenizer: {e}"))?;
+    let prompt = match &args.prompt_file {
+        Some(p) => std::fs::read_to_string(p).context("read prompt file")?,
+        None => "Explain, step by step, how a bicycle derailleur shifts gears.".into(),
+    };
+    let ids: Vec<u32> = tok
+        .encode(prompt.as_str(), false)
+        .map_err(|e| anyhow::anyhow!("tokenize: {e}"))?
+        .get_ids()
+        .to_vec();
+    anyhow::ensure!(ids.len() >= 2, "prompt is too short to prefill");
+    eprintln!("prompt: {} tokens", ids.len());
+
+    let sync = |d: &Device| -> Result<()> {
+        #[cfg(feature = "cuda")]
+        {
+            d.synchronize()?;
+        }
+        let _ = d;
+        Ok(())
+    };
+
+    // ── prefill ─────────────────────────────────────────────────────
+    let n = ids.len();
+    let t0 = Instant::now();
+    let (logits, hidden) = prefill(
+        &mut model,
+        &mut head,
+        &rotary,
+        &ids,
+        args.prefill_chunk,
+        dtype,
+        &dev,
+        &sync,
+    )?;
+    eprintln!("prefill ({n} tokens): {:.2} s", t0.elapsed().as_secs_f64());
+
+    // ── the loop ────────────────────────────────────────────────────
+    let vocab = text_cfg.vocab_size;
+    let mut last_logits: Vec<f32> = logits.i((0, 0, ..))?.to_dtype(DType::F32)?.to_vec1()?;
+    let mut hidden_cur = hidden;
+    let mut produced: Vec<u32> = Vec::new();
+    // (step at which it was drafted, the drafted tokens)
+    let mut pending: Vec<(usize, Vec<u32>)> = Vec::new();
+    let mut accepted_hist = vec![0usize; args.draft_len + 1];
+    let mut accepted_hist_penalised = vec![0usize; args.draft_len + 1];
+    let mut penalty_disagreements = 0usize;
+    let mut draft_ms_total = 0f64;
+    let mut draft_rounds = 0usize;
+    let mut target_ms_total = 0f64;
+    let mut target_steps = 0usize;
+
+    let total_rounds = args.warmup + args.steps;
+    for round in 0..total_rounds {
+        let measuring = round >= args.warmup;
+        let pos = n + produced.len(); // position of the token about to be drafted from
+        let next = argmax(&last_logits);
+        let next_penalised = argmax_with_penalty(
+            &last_logits,
+            &produced[produced.len().saturating_sub(args.repeat_last_n)..],
+            args.repeat_penalty,
+        );
+        if next != next_penalised {
+            penalty_disagreements += 1;
+        }
+        produced.push(next);
+
+        // Resolve any draft whose K true tokens have now been produced.
+        pending.retain(|(at, drafted)| {
+            let have = produced.len() - at;
+            if have < drafted.len() {
+                return true;
+            }
+            let truth = &produced[*at..*at + drafted.len()];
+            let acc = drafted
+                .iter()
+                .zip(truth.iter())
+                .take_while(|(d, t)| d == t)
+                .count();
+            accepted_hist[acc] += 1;
+            false
+        });
+
+        // ── draft, then rewind ──────────────────────────────────────
+        let snap = head.snapshot_kv()?;
+        let t0 = Instant::now();
+        let mut drafted = Vec::with_capacity(args.draft_len);
+        let mut cur_token = next;
+        let mut cur_hidden = hidden_cur.clone();
+        for k in 0..args.draft_len {
+            let tok_t = Tensor::new(&[cur_token], &dev)?.unsqueeze(0)?;
+            let emb = model.embed_tokens(&tok_t)?;
+            let (c, s) = rotary.cos_sin_at(&[pos + k])?;
+            let out = head.forward(&emb, &cur_hidden, None, &c, &s)?;
+            let row: Vec<f32> = model
+                .lm_head(&out)?
+                .i((0, 0, ..))?
+                .to_dtype(DType::F32)?
+                .to_vec1()?;
+            anyhow::ensure!(row.len() == vocab, "draft logits width {}", row.len());
+            cur_token = argmax(&row);
+            drafted.push(cur_token);
+            cur_hidden = out;
+        }
+        sync(&dev)?;
+        let draft_elapsed = t0.elapsed().as_secs_f64() * 1000.0;
+        head.restore_kv(&snap)?;
+
+        // The head must still advance over the *true* token, or its
+        // cache falls behind the target's.
+        let tok_t = Tensor::new(&[next], &dev)?.unsqueeze(0)?;
+        let emb = model.embed_tokens(&tok_t)?;
+        let (c, s) = rotary.cos_sin_at(&[pos])?;
+        head.forward(&emb, &hidden_cur, None, &c, &s)?;
+
+        // ── the target's own step ───────────────────────────────────
+        let t0 = Instant::now();
+        let (logits, h) = model.forward_with_hidden(&tok_t, pos)?;
+        sync(&dev)?;
+        let target_elapsed = t0.elapsed().as_secs_f64() * 1000.0;
+        last_logits = logits.i((0, 0, ..))?.to_dtype(DType::F32)?.to_vec1()?;
+        hidden_cur = h;
+
+        if measuring {
+            pending.push((produced.len(), drafted));
+            draft_ms_total += draft_elapsed;
+            draft_rounds += 1;
+            target_ms_total += target_elapsed;
+            target_steps += 1;
+        }
+    }
+
+    // Anything still pending never got its K true tokens; score what is
+    // known rather than dropping it silently.
+    for (at, drafted) in &pending {
+        let have = produced.len() - at;
+        let truth = &produced[*at..];
+        let acc = drafted
+            .iter()
+            .zip(truth.iter())
+            .take_while(|(d, t)| d == t)
+            .count();
+        if acc < have {
+            accepted_hist[acc] += 1;
+        }
+    }
+    let _ = &mut accepted_hist_penalised;
+
+    // ── verify cost, sampled ────────────────────────────────────────
+    //
+    // Timed after the loop because each sample advances the target's
+    // cache by K+1 positions — harmless once measuring is done, and it
+    // spares the loop the re-prefill that undoing one sample required.
+    let mut verify_samples = Vec::with_capacity(args.verify_samples);
+    for i in 0..args.verify_samples {
+        let block: Vec<u32> = std::iter::repeat_n(produced[0], args.draft_len + 1).collect();
+        let blk = Tensor::new(block.as_slice(), &dev)?.unsqueeze(0)?;
+        let at = n + produced.len() + i * (args.draft_len + 1);
+        let t0 = Instant::now();
+        let _ = model.forward_multi(&blk, at)?;
+        sync(&dev)?;
+        verify_samples.push(t0.elapsed().as_secs_f64() * 1000.0);
+    }
+    verify_samples.sort_by(f64::total_cmp);
+    let verify_ms = verify_samples[verify_samples.len() / 2];
+
+    // ── report ──────────────────────────────────────────────────────
+    let rounds: usize = accepted_hist.iter().sum();
+    let mean_accepted: f64 = accepted_hist
+        .iter()
+        .enumerate()
+        .map(|(k, n)| (k * n) as f64)
+        .sum::<f64>()
+        / rounds.max(1) as f64;
+    let target_ms = target_ms_total / target_steps.max(1) as f64;
+    let draft_ms = draft_ms_total / (draft_rounds.max(1) * args.draft_len) as f64;
+    let c = draft_ms / target_ms;
+    // One round commits 1 + accepted tokens and costs one verify plus K
+    // drafts, against (1 + accepted) plain decode steps.
+    let round_cost = verify_ms + args.draft_len as f64 * draft_ms;
+    let speedup = (1.0 + mean_accepted) * target_ms / round_cost;
+
+    let summary = serde_json::json!({
+        "model": args.model.to_string_lossy(),
+        "device": args.device,
+        "dtype": format!("{dtype:?}"),
+        "prompt_tokens": n,
+        "draft_len": args.draft_len,
+        "rounds_scored": rounds,
+        "target_decode_ms": target_ms,
+        "draft_step_ms": draft_ms,
+        "verify_ms_k_plus_1": verify_ms,
+        "verify_ms_min": verify_samples.first(),
+        "verify_ms_max": verify_samples.last(),
+        "verify_sample_count": verify_samples.len(),
+        "c_draft_over_target": c,
+        "mean_accepted_tokens": mean_accepted,
+        "accepted_histogram": accepted_hist,
+        "predicted_speedup": speedup,
+        "penalty_argmax_disagreements": penalty_disagreements,
+        "penalty_argmax_disagreement_rate":
+            penalty_disagreements as f64 / produced.len().max(1) as f64,
+    });
+    let pretty = serde_json::to_string_pretty(&summary)?;
+    println!("{pretty}");
+    if let Some(p) = &args.json {
+        std::fs::write(p, &pretty).context("write --json")?;
+    }
+    Ok(())
+}
+
+/// Tensor-parallel measurement (#96).
+///
+/// The only route to the 27B: 54 GB in bf16 does not fit one 5090, and
+/// quantisation is reachable only through the TP loader. The shape
+/// mirrors the single-GPU path — prefill, then draft-and-rewind at each
+/// decode step — with two differences:
+///
+/// - the target's step fans out to every rank, because the row-parallel
+///   `AllReduce`s only complete when all of them arrive;
+/// - the draft head is unsharded on the leader, so drafting issues no
+///   collectives at all and cannot leave a rank waiting.
+#[cfg(feature = "cuda")]
+async fn run_tp(args: Args) -> Result<()> {
+    use neuron::harness::device_worker::DeviceWorkerHandle;
+    use neuron::harness::tp::WorkerPool;
+
+    let cfg_raw =
+        std::fs::read_to_string(args.model.join("config.json")).context("read config.json")?;
+    let config: Config = Config::from_config_json(&cfg_raw).context("parse config.json")?;
+    let text_cfg = config.text_config.clone();
+    anyhow::ensure!(
+        MtpHead::present_in(&text_cfg),
+        "this checkpoint declares no MTP head (mtp_num_hidden_layers = 0)"
+    );
+
+    let mut shards: Vec<PathBuf> = std::fs::read_dir(&args.model)
+        .context("read model dir")?
+        .filter_map(|e| e.ok().map(|e| e.path()))
+        .filter(|p| p.extension().is_some_and(|x| x == "safetensors"))
+        .collect();
+    shards.sort();
+    anyhow::ensure!(!shards.is_empty(), "no .safetensors in {:?}", args.model);
+
+    let worker_bin = match &args.worker_binary {
+        Some(p) => p.clone(),
+        None => std::env::current_exe()
+            .context("resolve current_exe")?
+            .parent()
+            .context("executable has no parent directory")?
+            .join("neuron"),
+    };
+    anyhow::ensure!(
+        worker_bin.exists(),
+        "worker binary {worker_bin:?} not found — build `neuron` beside this probe, \
+         or pass --worker-binary"
+    );
+
+    let devices: Vec<u32> = (0..args.tensor_parallel).collect();
+    eprintln!(
+        "tp-{}: spawning workers from {worker_bin:?} on devices {devices:?}…",
+        args.tensor_parallel
+    );
+    let leader = DeviceWorkerHandle::spawn(devices[0])?;
+    let mut pool =
+        WorkerPool::spawn(&worker_bin, args.tensor_parallel, &devices, leader.clone()).await?;
+    pool.init_nccl(devices[0]).await?;
+
+    let leader_device = Device::new_cuda(devices[0] as usize)?;
+    let model_id = "mtp-probe";
+    let t0 = Instant::now();
+    let handle = pool
+        .load_dense_shard(
+            model_id,
+            &cfg_raw,
+            &shards,
+            &leader_device,
+            DType::BF16,
+            args.quant.clone(),
+        )
+        .await?;
+    eprintln!("shards loaded: {:.1} s", t0.elapsed().as_secs_f64());
+
+    let shard_strs: Vec<String> = shards
+        .iter()
+        .map(|p| p.to_string_lossy().into_owned())
+        .collect();
+    leader
+        .tp_load_mtp_head(handle, cfg_raw.clone(), shard_strs)
+        .await?;
+    eprintln!("draft head loaded on the leader");
+
+    let tok = tokenizers::Tokenizer::from_file(args.model.join("tokenizer.json"))
+        .map_err(|e| anyhow::anyhow!("load tokenizer: {e}"))?;
+    let prompt = match &args.prompt_file {
+        Some(p) => std::fs::read_to_string(p).context("read prompt file")?,
+        None => "Explain, step by step, how a bicycle derailleur shifts gears.".into(),
+    };
+    let ids: Vec<u32> = tok
+        .encode(prompt.as_str(), false)
+        .map_err(|e| anyhow::anyhow!("tokenize: {e}"))?
+        .get_ids()
+        .to_vec();
+    let n = ids.len();
+    anyhow::ensure!(n >= 2, "prompt is too short to prefill");
+    eprintln!("prompt: {n} tokens");
+
+    // ── prefill, chunked, target then head ──────────────────────────
+    let t0 = Instant::now();
+    let mut last_logits: Vec<f32> = Vec::new();
+    let mut at = 0usize;
+    while at < n {
+        let end = (at + args.prefill_chunk).min(n);
+        last_logits = pool
+            .generate_step(model_id, handle, ids[at..end].to_vec(), at)
+            .await?;
+        let head_end = end.min(n - 1);
+        if head_end > at {
+            leader
+                .tp_mtp_prefill_chunk(handle, ids[at + 1..head_end + 1].to_vec(), at)
+                .await?;
+        }
+        at = end;
+    }
+    eprintln!("prefill ({n} tokens): {:.1} s", t0.elapsed().as_secs_f64());
+
+    // ── draft / verify observation ──────────────────────────────────
+    let mut produced: Vec<u32> = Vec::new();
+    let mut pending: Vec<(usize, Vec<u32>)> = Vec::new();
+    let mut accepted_hist = vec![0usize; args.draft_len + 1];
+    let mut penalty_disagreements = 0usize;
+    let mut draft_ms_total = 0f64;
+    let mut target_ms_total = 0f64;
+    let mut measured = 0usize;
+
+    for round in 0..args.warmup + args.steps {
+        let measuring = round >= args.warmup;
+        let pos = n + produced.len();
+        let next = argmax(&last_logits);
+        let next_penalised = argmax_with_penalty(
+            &last_logits,
+            &produced[produced.len().saturating_sub(args.repeat_last_n)..],
+            args.repeat_penalty,
+        );
+        if next != next_penalised {
+            penalty_disagreements += 1;
+        }
+        produced.push(next);
+
+        pending.retain(|(start, drafted)| {
+            if produced.len() - start < drafted.len() {
+                return true;
+            }
+            let truth = &produced[*start..*start + drafted.len()];
+            let acc = drafted
+                .iter()
+                .zip(truth.iter())
+                .take_while(|(d, t)| d == t)
+                .count();
+            accepted_hist[acc] += 1;
+            false
+        });
+
+        let t0 = Instant::now();
+        let drafted = leader
+            .tp_mtp_draft(handle, next, pos, args.draft_len)
+            .await?;
+        let draft_elapsed = t0.elapsed().as_secs_f64() * 1000.0;
+        leader.tp_mtp_advance(handle, next, pos).await?;
+
+        let t0 = Instant::now();
+        last_logits = pool
+            .generate_step(model_id, handle, vec![next], pos)
+            .await?;
+        let target_elapsed = t0.elapsed().as_secs_f64() * 1000.0;
+
+        if measuring {
+            pending.push((produced.len(), drafted));
+            draft_ms_total += draft_elapsed;
+            target_ms_total += target_elapsed;
+            measured += 1;
+        }
+    }
+
+    // ── verify cost, sampled ────────────────────────────────────────
+    //
+    // After the loop, because each sample advances the target's cache by
+    // K+1. One sample was not enough: it moved 57.3 -> 50.1 ms between
+    // otherwise identical runs, wide enough to decide the gate by itself.
+    let mut verify_samples = Vec::with_capacity(args.verify_samples);
+    for i in 0..args.verify_samples {
+        let block: Vec<u32> = std::iter::repeat_n(produced[0], args.draft_len + 1).collect();
+        let at = n + produced.len() + i * (args.draft_len + 1);
+        let t0 = Instant::now();
+        let _ = pool
+            .generate_step_multi(model_id, handle, block, at)
+            .await?;
+        verify_samples.push(t0.elapsed().as_secs_f64() * 1000.0);
+    }
+    verify_samples.sort_by(f64::total_cmp);
+    let verify_ms = verify_samples[verify_samples.len() / 2];
+
+    let rounds: usize = accepted_hist.iter().sum();
+    let mean_accepted: f64 = accepted_hist
+        .iter()
+        .enumerate()
+        .map(|(k, c)| (k * c) as f64)
+        .sum::<f64>()
+        / rounds.max(1) as f64;
+    let target_ms = target_ms_total / measured.max(1) as f64;
+    let draft_ms = draft_ms_total / (measured.max(1) * args.draft_len) as f64;
+    let round_cost = verify_ms + args.draft_len as f64 * draft_ms;
+    let summary = serde_json::json!({
+        "model": args.model.to_string_lossy(),
+        "tensor_parallel": args.tensor_parallel,
+        "quant": args.quant,
+        "prompt_tokens": n,
+        "draft_len": args.draft_len,
+        "rounds_scored": rounds,
+        "target_decode_ms": target_ms,
+        "draft_step_ms": draft_ms,
+        "verify_ms_k_plus_1": verify_ms,
+        "verify_ms_min": verify_samples.first(),
+        "verify_ms_max": verify_samples.last(),
+        "verify_sample_count": verify_samples.len(),
+        "c_draft_over_target": draft_ms / target_ms,
+        "mean_accepted_tokens": mean_accepted,
+        "accepted_histogram": accepted_hist,
+        "predicted_speedup": (1.0 + mean_accepted) * target_ms / round_cost,
+        "penalty_argmax_disagreements": penalty_disagreements,
+        "penalty_argmax_disagreement_rate":
+            penalty_disagreements as f64 / produced.len().max(1) as f64,
+    });
+    let pretty = serde_json::to_string_pretty(&summary)?;
+    println!("{pretty}");
+    if let Some(p) = &args.json {
+        std::fs::write(p, &pretty).context("write --json")?;
+    }
+    pool.unload_model(model_id).await?;
+    Ok(())
+}
+
+#[cfg(not(feature = "cuda"))]
+async fn run_tp(_args: Args) -> Result<()> {
+    anyhow::bail!("--tensor-parallel > 1 requires a build with the `cuda` feature")
+}

crates/neuron/src/harness/arch/laya/backbone.rs

@@ -0,0 +1,478 @@
+//! ModernBERT, the bidirectional encoder under every Laya checkpoint.
+//!
+//! Follows the structure of `candle_transformers::models::modernbert`
+//! (itself a port of transformers' `ModernBertModel`), vendored rather
+//! than used as-is because that module only runs correctly in f32:
+//!
+//! - its attention masks are built in F32 whatever the model dtype, so
+//!   an f16/bf16 load fails at the first `broadcast_add`;
+//! - it computes the rotary angles `position × inv_freq` in the model
+//!   dtype, which in f16 already loses a quarter-radian of phase by
+//!   position 500;
+//! - it loads each layer's `attn_norm` with `.ok()`, so a weight-prefix
+//!   mistake silently yields a model with no pre-attention norms.
+//!
+//! Here the angles are computed in f32 and only the resulting cos/sin
+//! tables are cast; attention scores are softmaxed in f32 against f32
+//! masks; and `attn_norm` is required on every layer except layer 0,
+//! which ModernBERT builds as an identity.
+//!
+//! Precision follows PyTorch autocast, which is how the reference runs on
+//! a GPU: the linear layers (and so the attention inputs) run in the
+//! load dtype, while the residual stream, the embeddings and every
+//! LayerNorm stay in f32. Casting the residual stream to bf16 as well
+//! drifts the released English checkpoint's logits by ~0.25 over its 28
+//! layers; keeping it in f32 is what autocast does and costs little.
+//!
+//! Masked positions use a large *finite* negative rather than `-inf`.
+//! A padded query near the end of a short row can find every key in
+//! its sliding window padded; with `-inf` that row's softmax is NaN,
+//! and NaN in a padded row's value vector poisons the valid rows that
+//! give it zero weight (`0 × NaN = NaN`). A finite value makes such a
+//! row uniform instead, and valid rows never look at it.
+
+use anyhow::{Context, Result, ensure};
+use candle_core::{D, DType, Device, Module, Tensor};
+use candle_nn::{Embedding, LayerNorm, Linear, VarBuilder};
+
+use super::config::{EncoderConfig, LayerKind};
+
+/// Additive mask value for a position that must not be attended.
+pub(crate) const MASKED: f32 = -1e9;
+
+/// Apply `lin` in its weight's dtype, casting `x` in if needed. The
+/// result stays in that dtype; callers cast back to f32 where it joins
+/// the residual stream.
+pub(crate) fn project(lin: &Linear, x: &Tensor) -> candle_core::Result<Tensor> {
+    x.to_dtype(lin.weight().dtype())?.apply(lin)
+}
+
+/// A bias-free LayerNorm that takes candle's fused kernel.
+///
+/// `candle_nn::LayerNorm` dispatches to the fused `layer_norm` op only
+/// when it has a bias; without one it falls back to six elementwise
+/// kernels (mean, subtract, square, mean, divide, scale), which was a
+/// quarter of the encoder's GPU time. ModernBERT's norms have no bias,
+/// so they get an explicit zero one: same result, one kernel.
+fn fused_layer_norm_no_bias(size: usize, eps: f64, vb: VarBuilder) -> Result<LayerNorm> {
+    let weight = vb.get(size, "weight")?;
+    let bias = weight.zeros_like()?;
+    Ok(LayerNorm::new(weight, bias, eps))
+}
+
+/// Precomputed rotary tables for one base, `(max_pos, head_dim / 2)`.
+struct Rotary {
+    cos: Tensor,
+    sin: Tensor,
+}
+
+impl Rotary {
+    fn new(cfg: &EncoderConfig, theta: f64, dtype: DType, device: &Device) -> Result<Self> {
+        let dim = cfg.head_dim();
+        let inv_freq: Vec<f32> = (0..dim)
+            .step_by(2)
+            .map(|i| (1.0 / theta.powf(i as f64 / dim as f64)) as f32)
+            .collect();
+        let n = inv_freq.len();
+        let inv_freq = Tensor::from_vec(inv_freq, (1, n), device)?;
+        let max_pos = cfg.max_position_embeddings;
+        let pos = Tensor::arange(0u32, max_pos as u32, device)?
+            .to_dtype(DType::F32)?
+            .reshape((max_pos, 1))?;
+        let freqs = pos.matmul(&inv_freq)?;
+        Ok(Self {
+            cos: freqs.cos()?.to_dtype(dtype)?,
+            sin: freqs.sin()?.to_dtype(dtype)?,
+        })
+    }
+
+    fn apply(&self, q: &Tensor, k: &Tensor, seq_len: usize) -> Result<(Tensor, Tensor)> {
+        let cos = self.cos.narrow(0, 0, seq_len)?;
+        let sin = self.sin.narrow(0, 0, seq_len)?;
+        Ok((
+            candle_nn::rotary_emb::rope(&q.contiguous()?, &cos, &sin)?,
+            candle_nn::rotary_emb::rope(&k.contiguous()?, &cos, &sin)?,
+        ))
+    }
+}
+
+struct Attention {
+    wqkv: Linear,
+    wo: Linear,
+    n_heads: usize,
+    head_dim: usize,
+    kind: LayerKind,
+}
+
+impl Attention {
+    fn forward(&self, xs: &Tensor, rotary: &Rotary, mask: Option<&Tensor>) -> Result<Tensor> {
+        let (b, l, d) = xs.dims3()?;
+        let qkv = project(&self.wqkv, xs)?
+            .reshape((b, l, 3, self.n_heads, self.head_dim))?
+            .permute((2, 0, 3, 1, 4))?;
+        // Rotary in the projection dtype, as under autocast; the scores
+        // are then taken in f32 (see `scaled_attention`).
+        let q = qkv.get(0)?;
+        let k = qkv.get(1)?;
+        let v = qkv.get(2)?.contiguous()?;
+        let (q, k) = rotary.apply(&q, &k, l)?;
+        let ctx = scaled_attention(&q, &k, &v, mask, self.head_dim)?;
+        let ctx = ctx.transpose(1, 2)?.reshape((b, l, d))?;
+        Ok(ctx.apply(&self.wo)?.to_dtype(DType::F32)?)
+    }
+}
+
+/// A batch packed without padding for the flash-attention path: every
+/// row's tokens back to back, `(total, …)`. Linear layers then do no work
+/// for padding, and flash varlen keeps rows from attending to each other.
+#[cfg(feature = "flash-attn")]
+pub struct Packed {
+    /// `(b + 1,)` u32 cumulative row lengths, as flash varlen takes them.
+    pub cu_seqlens: Tensor,
+    /// Longest row.
+    pub max_len: usize,
+    /// `(total,)` u32: each token's position within its own row.
+    pub positions: Tensor,
+    /// Where each row starts in the packed order.
+    pub offsets: Vec<usize>,
+}
+
+#[cfg(feature = "flash-attn")]
+impl Packed {
+    pub fn new(lens: &[usize], device: &Device) -> Result<Self> {
+        let mut cu = Vec::with_capacity(lens.len() + 1);
+        let mut offsets = Vec::with_capacity(lens.len());
+        let mut positions = Vec::with_capacity(lens.iter().sum());
+        let mut total = 0u32;
+        cu.push(0u32);
+        for &n in lens {
+            offsets.push(total as usize);
+            positions.extend(0..n as u32);
+            total += n as u32;
+            cu.push(total);
+        }
+        let n = cu.len();
+        Ok(Self {
+            cu_seqlens: Tensor::from_vec(cu, n, device)?,
+            max_len: lens.iter().copied().max().unwrap_or(0),
+            positions: Tensor::from_vec(positions, total as usize, device)?,
+            offsets,
+        })
+    }
+}
+
+/// Bidirectional flash attention over a packed batch. `q`, `k`, `v` are
+/// `(total, h, dh)`; `half_window` limits each query to keys within that
+/// distance on either side (ModernBERT's sliding layers), `None` attends
+/// the whole row. Scores are accumulated in f32 inside the kernel.
+#[cfg(feature = "flash-attn")]
+pub(crate) fn flash_attention(
+    q: &Tensor,
+    k: &Tensor,
+    v: &Tensor,
+    packed: &Packed,
+    half_window: Option<usize>,
+) -> Result<Tensor> {
+    let head_dim = q.dim(2)?;
+    let scale = (head_dim as f64).powf(-0.5) as f32;
+    Ok(candle_flash_attn::flash_attn_varlen_windowed(
+        q,
+        k,
+        v,
+        &packed.cu_seqlens,
+        &packed.cu_seqlens,
+        packed.max_len,
+        packed.max_len,
+        scale,
+        half_window,
+        half_window,
+    )?)
+}
+
+/// Split a packed `(total, 3·d)` qkv projection into contiguous
+/// `(total, h, dh)` q, k and v.
+#[cfg(feature = "flash-attn")]
+pub(crate) fn split_qkv(
+    qkv: &Tensor,
+    n_heads: usize,
+    head_dim: usize,
+) -> Result<(Tensor, Tensor, Tensor)> {
+    let t = qkv.dim(0)?;
+    let qkv = qkv.reshape((t, 3, n_heads, head_dim))?;
+    let part = |i: usize| -> Result<Tensor> { Ok(qkv.narrow(1, i, 1)?.squeeze(1)?.contiguous()?) };
+    Ok((part(0)?, part(1)?, part(2)?))
+}
+
+#[cfg(feature = "flash-attn")]
+impl Rotary {
+    /// Rotate packed `(total, h, dh)` q and k by each token's position.
+    fn apply_packed(&self, q: &Tensor, k: &Tensor, positions: &Tensor) -> Result<(Tensor, Tensor)> {
+        let cos = self.cos.index_select(positions, 0)?;
+        let sin = self.sin.index_select(positions, 0)?;
+        let rot = |x: &Tensor| -> Result<Tensor> {
+            Ok(candle_nn::rotary_emb::rope_thd(&x.unsqueeze(0)?, &cos, &sin)?.squeeze(0)?)
+        };
+        Ok((rot(q)?, rot(k)?))
+    }
+}
+
+#[cfg(feature = "flash-attn")]
+impl Attention {
+    /// `xs` is the packed f32 residual stream `(total, d)`.
+    fn forward_packed(
+        &self,
+        xs: &Tensor,
+        rotary: &Rotary,
+        packed: &Packed,
+        half_window: usize,
+    ) -> Result<Tensor> {
+        let (t, d) = xs.dims2()?;
+        let (q, k, v) = split_qkv(&project(&self.wqkv, xs)?, self.n_heads, self.head_dim)?;
+        let (q, k) = rotary.apply_packed(&q, &k, &packed.positions)?;
+        let window = match self.kind {
+            LayerKind::Sliding => Some(half_window),
+            LayerKind::Global => None,
+        };
+        let ctx = flash_attention(&q, &k, &v, packed, window)?.reshape((t, d))?;
+        Ok(ctx.apply(&self.wo)?.to_dtype(DType::F32)?)
+    }
+}
+
+/// `softmax(q kᵀ / √d + mask) v`, scores and softmax in f32.
+///
+/// `q`, `k`, `v` are `(b, h, l, d)`; `mask` is an f32 additive mask
+/// broadcastable to `(b, h, l, l)`, or `None` when nothing is masked —
+/// which saves a bandwidth-bound pass over the `(b, h, l, l)` scores.
+pub(crate) fn scaled_attention(
+    q: &Tensor,
+    k: &Tensor,
+    v: &Tensor,
+    mask: Option<&Tensor>,
+    head_dim: usize,
+) -> Result<Tensor> {
+    let dtype = v.dtype();
+    let scale = (head_dim as f64).powf(-0.5);
+    // Scores in f32 end to end: a bf16 `q kᵀ` is rounded to 8 mantissa
+    // bits before the softmax ever sees it, which on this model's
+    // logit-sized scores is most of the drift from the reference's GPU
+    // path (whose SDPA kernels keep scores in f32).
+    // The scale goes on q, not on the `(b, h, l, l)` scores: the same
+    // product, on a tensor `l / d` times smaller.
+    let scores = (q.to_dtype(DType::F32)? * scale)?.matmul(&k.to_dtype(DType::F32)?.t()?)?;
+    let scores = match mask {
+        Some(mask) => scores.broadcast_add(mask)?,
+        None => scores,
+    };
+    let probs = candle_nn::ops::softmax_last_dim(&scores)?;
+    Ok(probs.to_dtype(dtype)?.matmul(v)?)
+}
+
+struct Mlp {
+    wi: Linear,
+    wo: Linear,
+}
+
+impl Module for Mlp {
+    /// f32 in, f32 out; the projections run in the weights' dtype.
+    fn forward(&self, xs: &Tensor) -> candle_core::Result<Tensor> {
+        let xs = project(&self.wi, xs)?;
+        let parts = xs.chunk(2, D::Minus1)?;
+        (parts[0].gelu_erf()? * &parts[1])?
+            .apply(&self.wo)?
+            .to_dtype(DType::F32)
+    }
+}
+
+struct Layer {
+    attn_norm: Option<LayerNorm>,
+    attn: Attention,
+    mlp_norm: LayerNorm,
+    mlp: Mlp,
+}
+
+/// The two additive masks one forward needs, both f32. Either is `None`
+/// when it would mask nothing: `global` for a batch with no padding (every
+/// single-question request), `sliding` additionally when no row is longer
+/// than the window reaches.
+pub struct Masks {
+    /// `(b, 1, 1, l)`: padded keys masked. Global layers and the
+    /// decision head use this one.
+    pub global: Option<Tensor>,
+    /// `(b, 1, l, l)`: padded keys, plus keys outside the window.
+    pub sliding: Option<Tensor>,
+}
+
+impl Masks {
+    /// Build both masks from per-row valid lengths. Rows are
+    /// right-padded: positions `>= lens[i]` of row `i` are padding.
+    pub fn new(
+        lens: &[usize],
+        seq_len: usize,
+        half_window: usize,
+        device: &Device,
+    ) -> Result<Self> {
+        let b = lens.len();
+        let padded = lens.iter().any(|&n| n < seq_len);
+        let windowed = seq_len > half_window + 1;
+        let global = if padded {
+            let mut pad = Vec::with_capacity(b * seq_len);
+            for &n in lens {
+                pad.extend((0..seq_len).map(|j| if j < n { 0f32 } else { MASKED }));
+            }
+            Some(Tensor::from_vec(pad, (b, 1, 1, seq_len), device)?)
+        } else {
+            None
+        };
+        if !windowed {
+            let sliding = global.clone();
+            return Ok(Self { global, sliding });
+        }
+        // Built on the device from two aranges, not on the host: at
+        // 1024 tokens a host-built window is a 4 MB upload per forward.
+        let pos = Tensor::arange(0u32, seq_len as u32, device)?.to_dtype(DType::F32)?;
+        let dist = pos
+            .reshape((seq_len, 1))?
+            .broadcast_sub(&pos.reshape((1, seq_len))?)?
+            .abs()?;
+        let outside = dist.gt(half_window as f64)?;
+        let zeros = Tensor::zeros((seq_len, seq_len), DType::F32, device)?;
+        let masked = Tensor::full(MASKED, (seq_len, seq_len), device)?;
+        let window = outside
+            .where_cond(&masked, &zeros)?
+            .reshape((1, 1, seq_len, seq_len))?;
+        let sliding = match &global {
+            Some(g) => g.broadcast_add(&window)?,
+            None => window,
+        };
+        Ok(Self {
+            global,
+            sliding: Some(sliding),
+        })
+    }
+}
+
+pub struct ModernBert {
+    embeddings: Embedding,
+    embed_norm: LayerNorm,
+    layers: Vec<Layer>,
+    final_norm: LayerNorm,
+    global_rotary: Rotary,
+    sliding_rotary: Rotary,
+    half_window: usize,
+}
+
+impl ModernBert {
+    /// Load from a VarBuilder rooted at the encoder's prefix (`encoder`
+    /// in a Laya checkpoint, `model` in a stock ModernBERT one). Linear
+    /// weights load in the VarBuilder's dtype; embeddings and norms in
+    /// f32.
+    pub fn load(vb: VarBuilder, cfg: &EncoderConfig) -> Result<Self> {
+        let d = cfg.hidden_size;
+        let f32_vb = vb.clone().set_dtype(DType::F32);
+        let embeddings = candle_nn::embedding(
+            cfg.vocab_size,
+            d,
+            f32_vb.pp("embeddings").pp("tok_embeddings"),
+        )
+        .context("encoder.embeddings.tok_embeddings")?;
+        let embed_norm = fused_layer_norm_no_bias(d, cfg.norm_eps, f32_vb.pp("embeddings.norm"))?;
+        let mut layers = Vec::with_capacity(cfg.num_hidden_layers);
+        for (i, kind) in cfg.layer_kinds.iter().copied().enumerate() {
+            let lvb = vb.pp(format!("layers.{i}"));
+            let norm_vb = f32_vb.pp(format!("layers.{i}"));
+            // ModernBERT builds layer 0's attn_norm as an identity and
+            // saves nothing for it; every other layer must have one.
+            let attn_norm = if lvb.contains_tensor("attn_norm.weight") {
+                Some(fused_layer_norm_no_bias(
+                    d,
+                    cfg.norm_eps,
+                    norm_vb.pp("attn_norm"),
+                )?)
+            } else {
+                ensure!(
+                    i == 0,
+                    "encoder layer {i} has no attn_norm.weight — wrong weight prefix, \
+                     or not a ModernBERT checkpoint"
+                );
+                None
+            };
+            let attn = Attention {
+                wqkv: candle_nn::linear_no_bias(d, 3 * d, lvb.pp("attn.Wqkv"))?,
+                wo: candle_nn::linear_no_bias(d, d, lvb.pp("attn.Wo"))?,
+                n_heads: cfg.num_attention_heads,
+                head_dim: cfg.head_dim(),
+                kind,
+            };
+            let mlp_norm = fused_layer_norm_no_bias(d, cfg.norm_eps, norm_vb.pp("mlp_norm"))?;
+            let mlp = Mlp {
+                wi: candle_nn::linear_no_bias(d, 2 * cfg.intermediate_size, lvb.pp("mlp.Wi"))?,
+                wo: candle_nn::linear_no_bias(cfg.intermediate_size, d, lvb.pp("mlp.Wo"))?,
+            };
+            layers.push(Layer {
+                attn_norm,
+                attn,
+                mlp_norm,
+                mlp,
+            });
+        }
+        let final_norm = fused_layer_norm_no_bias(d, cfg.norm_eps, f32_vb.pp("final_norm"))?;
+        let device = vb.device();
+        Ok(Self {
+            embeddings,
+            embed_norm,
+            layers,
+            final_norm,
+            global_rotary: Rotary::new(cfg, cfg.global_rope_theta, vb.dtype(), device)?,
+            sliding_rotary: Rotary::new(cfg, cfg.local_rope_theta, vb.dtype(), device)?,
+            half_window: cfg.local_attention / 2,
+        })
+    }
+
+    pub fn half_window(&self) -> usize {
+        self.half_window
+    }
+
+    /// `ids` is `(b, l)` u32; returns the final hidden states `(b, l, d)`
+    /// in f32.
+    pub fn forward(&self, ids: &Tensor, masks: &Masks) -> Result<Tensor> {
+        let mut xs = ids.apply(&self.embeddings)?.apply(&self.embed_norm)?;
+        for layer in &self.layers {
+            let normed = match &layer.attn_norm {
+                Some(norm) => xs.apply(norm)?,
+                None => xs.clone(),
+            };
+            let (rotary, mask) = match layer.attn.kind {
+                LayerKind::Global => (&self.global_rotary, masks.global.as_ref()),
+                LayerKind::Sliding => (&self.sliding_rotary, masks.sliding.as_ref()),
+            };
+            xs = (&xs + layer.attn.forward(&normed, rotary, mask)?)?;
+            let mlp_out = xs.apply(&layer.mlp_norm)?.apply(&layer.mlp)?;
+            xs = (xs + mlp_out)?;
+        }
+        Ok(xs.apply(&self.final_norm)?)
+    }
+
+    /// The flash path: `ids` is the packed `(total,)` u32 token stream;
+    /// returns the packed final hidden states `(total, d)` in f32.
+    #[cfg(feature = "flash-attn")]
+    pub fn forward_packed(&self, ids: &Tensor, packed: &Packed) -> Result<Tensor> {
+        let mut xs = ids.apply(&self.embeddings)?.apply(&self.embed_norm)?;
+        for layer in &self.layers {
+            let normed = match &layer.attn_norm {
+                Some(norm) => xs.apply(norm)?,
+                None => xs.clone(),
+            };
+            let rotary = match layer.attn.kind {
+                LayerKind::Global => &self.global_rotary,
+                LayerKind::Sliding => &self.sliding_rotary,
+            };
+            xs = (&xs
+                + layer
+                    .attn
+                    .forward_packed(&normed, rotary, packed, self.half_window)?)?;
+            let mlp_out = xs.apply(&layer.mlp_norm)?.apply(&layer.mlp)?;
+            xs = (xs + mlp_out)?;
+        }
+        Ok(xs.apply(&self.final_norm)?)
+    }
+}

crates/neuron/src/harness/arch/laya/config.rs

@@ -0,0 +1,358 @@
+//! The two config files a Laya checkpoint ships, and what they mean.
+//!
+//! A checkpoint directory holds `rl_agent_config.json` (the decision
+//! model: head depth, token budgets, calibration temperatures) and
+//! `encoder/config.json` (the ModernBERT backbone, written by
+//! transformers 5). There is no top-level `config.json`.
+//!
+//! The encoder config is parsed here rather than through
+//! `candle_transformers::models::modernbert::Config` because
+//! transformers 5 moved the rotary bases into a nested
+//! `rope_parameters.{full_attention,sliding_attention}.rope_theta`
+//! block, which that struct cannot read; the older flat
+//! `global_rope_theta` / `local_rope_theta` fields are still honoured
+//! when a checkpoint carries them. mmBERT (the multilingual
+//! checkpoint's backbone) is the case that matters: both of its bases
+//! are 160,000, where the flat-field defaults would silently give the
+//! sliding layers 10,000.
+
+use anyhow::{Context, Result, bail, ensure};
+use serde::Deserialize;
+use std::collections::BTreeMap;
+use std::path::Path;
+
+/// `rl_agent_config.json`.
+#[derive(Debug, Clone, Deserialize)]
+pub struct LayaConfig {
+    /// The backbone the head was trained on, e.g.
+    /// `answerdotai/ModernBERT-large`. Informational: the weights come
+    /// from the checkpoint, the shape from `encoder/config.json`.
+    #[serde(default)]
+    pub encoder: Option<String>,
+    #[serde(default = "default_head_layers")]
+    pub head_layers: usize,
+    #[serde(default = "default_max_len")]
+    pub max_len: usize,
+    #[serde(default = "default_head_max_len")]
+    pub head_max_len: usize,
+    /// One act head output per entry, plus one for "answer". The act
+    /// head's width is `act_costs.len() + 1`.
+    #[serde(default)]
+    pub act_costs: BTreeMap<String, f64>,
+    /// Per-question-type temperatures, indexed choice / score / noul.
+    #[serde(default = "default_temperature")]
+    pub temperature: Vec<f64>,
+    /// Temperatures fitted per `"<type>:<cardinality bucket>"`.
+    #[serde(default)]
+    pub temperature_by_options: BTreeMap<String, f64>,
+    /// Autocast dtype the reference runs on GPU (`"bf16"` or `"fp16"`).
+    #[serde(default)]
+    pub amp_dtype: Option<String>,
+}
+
+fn default_head_layers() -> usize {
+    2
+}
+fn default_max_len() -> usize {
+    512
+}
+fn default_head_max_len() -> usize {
+    192
+}
+fn default_temperature() -> Vec<f64> {
+    vec![1.0, 1.0, 1.0]
+}
+
+impl LayaConfig {
+    pub fn n_act(&self) -> usize {
+        self.act_costs.len() + 1
+    }
+}
+
+/// Whether an attention layer sees the whole sequence or a window.
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub enum LayerKind {
+    Global,
+    Sliding,
+}
+
+/// The ModernBERT backbone shape, resolved from `encoder/config.json`.
+#[derive(Debug, Clone)]
+pub struct EncoderConfig {
+    pub vocab_size: usize,
+    pub hidden_size: usize,
+    pub num_hidden_layers: usize,
+    pub num_attention_heads: usize,
+    pub intermediate_size: usize,
+    pub max_position_embeddings: usize,
+    pub norm_eps: f64,
+    pub pad_token_id: u32,
+    /// Full sliding-window width. A token attends to keys at distance
+    /// `<= local_attention / 2` on either side.
+    pub local_attention: usize,
+    pub global_rope_theta: f64,
+    pub local_rope_theta: f64,
+    pub layer_kinds: Vec<LayerKind>,
+}
+
+#[derive(Deserialize)]
+struct RawEncoderConfig {
+    vocab_size: usize,
+    hidden_size: usize,
+    num_hidden_layers: usize,
+    num_attention_heads: usize,
+    intermediate_size: usize,
+    max_position_embeddings: usize,
+    #[serde(default)]
+    norm_eps: Option<f64>,
+    #[serde(default)]
+    layer_norm_eps: Option<f64>,
+    pad_token_id: u32,
+    #[serde(default = "default_global_every")]
+    global_attn_every_n_layers: usize,
+    local_attention: usize,
+    #[serde(default)]
+    layer_types: Option<Vec<String>>,
+    #[serde(default)]
+    rope_parameters: Option<serde_json::Value>,
+    #[serde(default)]
+    global_rope_theta: Option<f64>,
+    #[serde(default)]
+    local_rope_theta: Option<f64>,
+    #[serde(default)]
+    norm_bias: bool,
+    #[serde(default)]
+    attention_bias: bool,
+    #[serde(default)]
+    mlp_bias: bool,
+    #[serde(default)]
+    hidden_activation: Option<String>,
+}
+
+fn default_global_every() -> usize {
+    3
+}
+
+/// transformers' ModernBERT defaults for checkpoints that predate
+/// `rope_parameters` and also omit the flat fields.
+const DEFAULT_GLOBAL_ROPE_THETA: f64 = 160_000.0;
+const DEFAULT_LOCAL_ROPE_THETA: f64 = 10_000.0;
+
+impl EncoderConfig {
+    pub fn from_json(text: &str) -> Result<Self> {
+        let raw: RawEncoderConfig =
+            serde_json::from_str(text).context("parse ModernBERT encoder config")?;
+
+        // Variants this port does not implement are refused, not
+        // approximated: each would load without error and answer wrong.
+        ensure!(
+            !raw.norm_bias && !raw.attention_bias && !raw.mlp_bias,
+            "ModernBERT with norm/attention/mlp biases is not supported"
+        );
+        if let Some(act) = raw.hidden_activation.as_deref() {
+            ensure!(
+                act == "gelu",
+                "ModernBERT hidden_activation {act:?} is not supported (only exact \"gelu\")"
+            );
+        }
+        ensure!(
+            raw.hidden_size.is_multiple_of(raw.num_attention_heads),
+            "hidden_size {} is not divisible by num_attention_heads {}",
+            raw.hidden_size,
+            raw.num_attention_heads
+        );
+
+        let (global_rope_theta, local_rope_theta) = rope_thetas(&raw)?;
+
+        let layer_kinds = match &raw.layer_types {
+            Some(types) => {
+                ensure!(
+                    types.len() == raw.num_hidden_layers,
+                    "layer_types has {} entries for {} layers",
+                    types.len(),
+                    raw.num_hidden_layers
+                );
+                types
+                    .iter()
+                    .map(|t| match t.as_str() {
+                        "full_attention" => Ok(LayerKind::Global),
+                        "sliding_attention" => Ok(LayerKind::Sliding),
+                        other => bail!("unknown ModernBERT layer type {other:?}"),
+                    })
+                    .collect::<Result<Vec<_>>>()?
+            }
+            None => (0..raw.num_hidden_layers)
+                .map(|i| {
+                    if i % raw.global_attn_every_n_layers == 0 {
+                        LayerKind::Global
+                    } else {
+                        LayerKind::Sliding
+                    }
+                })
+                .collect(),
+        };
+
+        Ok(Self {
+            vocab_size: raw.vocab_size,
+            hidden_size: raw.hidden_size,
+            num_hidden_layers: raw.num_hidden_layers,
+            num_attention_heads: raw.num_attention_heads,
+            intermediate_size: raw.intermediate_size,
+            max_position_embeddings: raw.max_position_embeddings,
+            norm_eps: raw.norm_eps.or(raw.layer_norm_eps).unwrap_or(1e-5),
+            pad_token_id: raw.pad_token_id,
+            local_attention: raw.local_attention,
+            global_rope_theta,
+            local_rope_theta,
+            layer_kinds,
+        })
+    }
+
+    pub fn head_dim(&self) -> usize {
+        self.hidden_size / self.num_attention_heads
+    }
+}
+
+/// Resolve the two rotary bases. `rope_parameters` (transformers 5)
+/// wins over the flat fields; within it, a per-layer-type block wins
+/// over a shared top-level `rope_theta`.
+fn rope_thetas(raw: &RawEncoderConfig) -> Result<(f64, f64)> {
+    let mut global = raw.global_rope_theta.unwrap_or(DEFAULT_GLOBAL_ROPE_THETA);
+    let mut local = raw.local_rope_theta.unwrap_or(DEFAULT_LOCAL_ROPE_THETA);
+    if let Some(rope) = &raw.rope_parameters {
+        let shared = rope.get("rope_theta").and_then(|v| v.as_f64());
+        for (key, slot) in [
+            ("full_attention", &mut global),
+            ("sliding_attention", &mut local),
+        ] {
+            let block = rope.get(key);
+            if let Some(rope_type) = block
+                .and_then(|b| b.get("rope_type"))
+                .and_then(|v| v.as_str())
+            {
+                ensure!(
+                    rope_type == "default",
+                    "rope_type {rope_type:?} for {key} is not supported"
+                );
+            }
+            let theta = block
+                .and_then(|b| b.get("rope_theta"))
+                .and_then(|v| v.as_f64())
+                .or(shared);
+            if let Some(theta) = theta {
+                *slot = theta;
+            }
+        }
+    }
+    Ok((global, local))
+}
+
+/// Both configs of one checkpoint directory.
+#[derive(Debug, Clone)]
+pub struct CheckpointConfig {
+    pub laya: LayaConfig,
+    pub encoder: EncoderConfig,
+}
+
+impl CheckpointConfig {
+    /// Read `rl_agent_config.json` and `encoder/config.json` from a
+    /// checkpoint directory (the repo root, or one of its subfolders).
+    pub fn from_dir(dir: &Path) -> Result<Self> {
+        let laya_path = dir.join("rl_agent_config.json");
+        let laya: LayaConfig = serde_json::from_str(
+            &std::fs::read_to_string(&laya_path)
+                .with_context(|| format!("read {}", laya_path.display()))?,
+        )
+        .with_context(|| format!("parse {}", laya_path.display()))?;
+        let enc_path = dir.join("encoder").join("config.json");
+        let encoder = EncoderConfig::from_json(
+            &std::fs::read_to_string(&enc_path)
+                .with_context(|| format!("read {}", enc_path.display()))?,
+        )
+        .with_context(|| format!("parse {}", enc_path.display()))?;
+        ensure!(
+            laya.temperature.len() == 3,
+            "rl_agent_config temperature must have 3 entries (choice, score, noul), got {}",
+            laya.temperature.len()
+        );
+        Ok(Self { laya, encoder })
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn base() -> serde_json::Value {
+        serde_json::json!({
+            "vocab_size": 100, "hidden_size": 64, "num_hidden_layers": 4,
+            "num_attention_heads": 4, "intermediate_size": 96,
+            "max_position_embeddings": 512, "norm_eps": 1e-5, "pad_token_id": 0,
+            "global_attn_every_n_layers": 3, "local_attention": 16,
+        })
+    }
+
+    #[test]
+    fn nested_rope_parameters_override_flat_defaults() {
+        let mut v = base();
+        v["rope_parameters"] = serde_json::json!({
+            "full_attention": {"rope_theta": 160000.0, "rope_type": "default"},
+            "sliding_attention": {"rope_theta": 160000.0, "rope_type": "default"},
+        });
+        let c = EncoderConfig::from_json(&v.to_string()).unwrap();
+        assert_eq!(c.global_rope_theta, 160_000.0);
+        // The mmBERT case: the flat default would have been 10,000.
+        assert_eq!(c.local_rope_theta, 160_000.0);
+    }
+
+    #[test]
+    fn flat_fields_are_read_without_rope_parameters() {
+        let mut v = base();
+        v["global_rope_theta"] = serde_json::json!(80000.0);
+        v["local_rope_theta"] = serde_json::json!(5000.0);
+        let c = EncoderConfig::from_json(&v.to_string()).unwrap();
+        assert_eq!((c.global_rope_theta, c.local_rope_theta), (80000.0, 5000.0));
+    }
+
+    #[test]
+    fn layer_kinds_follow_every_n_when_layer_types_absent() {
+        let c = EncoderConfig::from_json(&base().to_string()).unwrap();
+        assert_eq!(
+            c.layer_kinds,
+            vec![
+                LayerKind::Global,
+                LayerKind::Sliding,
+                LayerKind::Sliding,
+                LayerKind::Global
+            ]
+        );
+    }
+
+    #[test]
+    fn explicit_layer_types_win() {
+        let mut v = base();
+        v["layer_types"] = serde_json::json!([
+            "sliding_attention",
+            "full_attention",
+            "full_attention",
+            "sliding_attention"
+        ]);
+        let c = EncoderConfig::from_json(&v.to_string()).unwrap();
+        assert_eq!(c.layer_kinds[0], LayerKind::Sliding);
+        assert_eq!(c.layer_kinds[1], LayerKind::Global);
+    }
+
+    #[test]
+    fn unsupported_variants_are_refused() {
+        let mut v = base();
+        v["attention_bias"] = serde_json::json!(true);
+        assert!(EncoderConfig::from_json(&v.to_string()).is_err());
+        let mut v = base();
+        v["hidden_activation"] = serde_json::json!("gelu_pytorch_tanh");
+        assert!(EncoderConfig::from_json(&v.to_string()).is_err());
+        let mut v = base();
+        v["rope_parameters"] =
+            serde_json::json!({"full_attention": {"rope_type": "yarn", "rope_theta": 1.0}});
+        assert!(EncoderConfig::from_json(&v.to_string()).is_err());
+    }
+}

crates/neuron/src/harness/arch/laya/head.rs

@@ -0,0 +1,180 @@
+//! The decision head that sits on top of the encoder.
+//!
+//! Mirrors the reference `DecisionModel` (the `laya` SDK's
+//! `common.py`):
+//!
+//! 1. a per-question-type embedding added to every position;
+//! 2. `head_layers` × PyTorch `nn.TransformerEncoderLayer` with
+//!    `norm_first=True`: pre-norm self-attention through a fused
+//!    `in_proj` with bias, then a pre-norm feed-forward whose
+//!    activation is **ReLU** — PyTorch's default, which the reference
+//!    never overrides;
+//! 3. `scorer`, applied at each option's `[MASK]` marker: LayerNorm →
+//!    Linear → exact GELU → Linear to one logit;
+//! 4. `act_head`, over the first position's hidden state plus four
+//!    summary features of the (uncalibrated) answer distribution.
+
+use anyhow::Result;
+use candle_core::{DType, Tensor};
+use candle_nn::{Embedding, LayerNorm, Linear, VarBuilder};
+
+use super::backbone::{project, scaled_attention};
+
+/// PyTorch's `LayerNorm` default.
+const TORCH_LN_EPS: f64 = 1e-5;
+
+struct HeadLayer {
+    norm1: LayerNorm,
+    in_proj: Linear,
+    out_proj: Linear,
+    norm2: LayerNorm,
+    linear1: Linear,
+    linear2: Linear,
+    n_heads: usize,
+    head_dim: usize,
+}
+
+impl HeadLayer {
+    fn load(vb: VarBuilder, d: usize) -> Result<Self> {
+        // `nn.TransformerEncoder` is built with `nhead = max(1, d // 64)`
+        // and `dim_feedforward = 4 * d`.
+        let n_heads = (d / 64).max(1);
+        let f32_vb = vb.clone().set_dtype(DType::F32);
+        Ok(Self {
+            norm1: candle_nn::layer_norm(d, TORCH_LN_EPS, f32_vb.pp("norm1"))?,
+            in_proj: linear_torch_mha_in_proj(d, vb.pp("self_attn"))?,
+            out_proj: candle_nn::linear(d, d, vb.pp("self_attn.out_proj"))?,
+            norm2: candle_nn::layer_norm(d, TORCH_LN_EPS, f32_vb.pp("norm2"))?,
+            linear1: candle_nn::linear(d, 4 * d, vb.pp("linear1"))?,
+            linear2: candle_nn::linear(4 * d, d, vb.pp("linear2"))?,
+            n_heads,
+            head_dim: d / n_heads,
+        })
+    }
+
+    /// `xs` is the f32 residual stream; projections run in the weights'
+    /// dtype and join it back in f32, as under autocast.
+    fn forward(&self, xs: &Tensor, key_mask: Option<&Tensor>) -> Result<Tensor> {
+        let (b, l, d) = xs.dims3()?;
+        let qkv = project(&self.in_proj, &xs.apply(&self.norm1)?)?
+            .reshape((b, l, 3, self.n_heads, self.head_dim))?
+            .permute((2, 0, 3, 1, 4))?;
+        let q = qkv.get(0)?.contiguous()?;
+        let k = qkv.get(1)?.contiguous()?;
+        let v = qkv.get(2)?.contiguous()?;
+        let ctx = scaled_attention(&q, &k, &v, key_mask, self.head_dim)?
+            .transpose(1, 2)?
+            .reshape((b, l, d))?;
+        let xs = (xs + ctx.apply(&self.out_proj)?.to_dtype(DType::F32)?)?;
+        let ff = project(&self.linear1, &xs.apply(&self.norm2)?)?
+            .relu()?
+            .apply(&self.linear2)?
+            .to_dtype(DType::F32)?;
+        Ok((xs + ff)?)
+    }
+
+    /// The flash path over a packed `(total, d)` f32 stream.
+    #[cfg(feature = "flash-attn")]
+    fn forward_packed(&self, xs: &Tensor, packed: &super::backbone::Packed) -> Result<Tensor> {
+        use super::backbone::{flash_attention, split_qkv};
+        let (t, d) = xs.dims2()?;
+        let qkv = project(&self.in_proj, &xs.apply(&self.norm1)?)?;
+        let (q, k, v) = split_qkv(&qkv, self.n_heads, self.head_dim)?;
+        let ctx = flash_attention(&q, &k, &v, packed, None)?.reshape((t, d))?;
+        let xs = (xs + ctx.apply(&self.out_proj)?.to_dtype(DType::F32)?)?;
+        let ff = project(&self.linear1, &xs.apply(&self.norm2)?)?
+            .relu()?
+            .apply(&self.linear2)?
+            .to_dtype(DType::F32)?;
+        Ok((xs + ff)?)
+    }
+}
+
+/// `nn.MultiheadAttention` stores q/k/v as one `in_proj_weight`
+/// `(3d, d)` and `in_proj_bias` `(3d)`, rows ordered q then k then v —
+/// exactly the layout one fused `Linear` wants.
+fn linear_torch_mha_in_proj(d: usize, vb: VarBuilder) -> Result<Linear> {
+    let w = vb.get((3 * d, d), "in_proj_weight")?;
+    let b = vb.get(3 * d, "in_proj_bias")?;
+    Ok(Linear::new(w, Some(b)))
+}
+
+pub struct DecisionHead {
+    type_emb: Embedding,
+    layers: Vec<HeadLayer>,
+    scorer_norm: LayerNorm,
+    scorer_fc: Linear,
+    scorer_out: Linear,
+    act_fc: Linear,
+    act_out: Linear,
+}
+
+impl DecisionHead {
+    /// `vb` is the checkpoint root; the head's tensors are top-level.
+    /// Linear weights load in `vb`'s dtype; the type embedding and norms
+    /// in f32.
+    pub fn load(vb: VarBuilder, d: usize, head_layers: usize, n_act: usize) -> Result<Self> {
+        let layers = (0..head_layers)
+            .map(|i| HeadLayer::load(vb.pp(format!("head.layers.{i}")), d))
+            .collect::<Result<Vec<_>>>()?;
+        let f32_vb = vb.clone().set_dtype(DType::F32);
+        Ok(Self {
+            type_emb: candle_nn::embedding(3, d, f32_vb.pp("type_emb"))?,
+            layers,
+            scorer_norm: candle_nn::layer_norm(d, TORCH_LN_EPS, f32_vb.pp("scorer.0"))?,
+            scorer_fc: candle_nn::linear(d, d, vb.pp("scorer.1"))?,
+            scorer_out: candle_nn::linear(d, 1, vb.pp("scorer.3"))?,
+            act_fc: candle_nn::linear(d + 4, 256, vb.pp("act_head.0"))?,
+            act_out: candle_nn::linear(256, n_act, vb.pp("act_head.2"))?,
+        })
+    }
+
+    /// Type embedding plus the transformer layers. `h` is the encoder
+    /// output `(b, l, d)`, `qtype` is `(b,)` u32.
+    pub fn contextualise(
+        &self,
+        h: &Tensor,
+        qtype: &Tensor,
+        key_mask: Option<&Tensor>,
+    ) -> Result<Tensor> {
+        let mut h = h.broadcast_add(&qtype.apply(&self.type_emb)?.unsqueeze(1)?)?;
+        for layer in &self.layers {
+            h = layer.forward(&h, key_mask)?;
+        }
+        Ok(h)
+    }
+
+    /// The flash path: `h` is the packed encoder output `(total, d)` and
+    /// `qtype` the per-token question type `(total,)` u32.
+    #[cfg(feature = "flash-attn")]
+    pub fn contextualise_packed(
+        &self,
+        h: &Tensor,
+        qtype: &Tensor,
+        packed: &super::backbone::Packed,
+    ) -> Result<Tensor> {
+        let mut h = (h + qtype.apply(&self.type_emb)?)?;
+        for layer in &self.layers {
+            h = layer.forward_packed(&h, packed)?;
+        }
+        Ok(h)
+    }
+
+    /// Score gathered marker states `(b, k, d)` → `(b, k)` f32 logits.
+    pub fn score(&self, markers: &Tensor) -> Result<Tensor> {
+        let x = project(&self.scorer_fc, &markers.apply(&self.scorer_norm)?)?
+            .gelu_erf()?
+            .apply(&self.scorer_out)?;
+        Ok(x.squeeze(2)?.to_dtype(DType::F32)?)
+    }
+
+    /// Act logits `(b, n_act)` f32 from the pooled first position
+    /// `(b, d)` and the four answer-distribution features `(b, 4)`.
+    pub fn act(&self, pooled: &Tensor, feats: &Tensor) -> Result<Tensor> {
+        let x = Tensor::cat(&[pooled, &feats.to_dtype(pooled.dtype())?], 1)?;
+        let x = project(&self.act_fc, &x)?
+            .gelu_erf()?
+            .apply(&self.act_out)?;
+        Ok(x.to_dtype(DType::F32)?)
+    }
+}

crates/neuron/src/harness/arch/laya/mod.rs

@@ -0,0 +1,64 @@
+//! Laya — a non-autoregressive decision model (`convaiinnovations/laya`).
+//!
+//! Not a language model: a ModernBERT encoder plus a small decision
+//! head that scores the options of a typed question in one forward
+//! pass. The caller assembles each question into one token sequence
+//! (`[CLS] <type> question: <instructions> [SEP] [MASK] opt0 [MASK]
+//! opt1 … [SEP] state [SEP]`) and says where each option's `[MASK]`
+//! marker sits; this module returns one raw score per marker plus the
+//! act head's distribution. Calibration (temperature scaling) and
+//! answer shaping happen after, CPU-side, from the checkpoint's
+//! `rl_agent_config.json`.
+//!
+//! The reference implementation is the `laya` SDK's `DecisionModel`
+//! (`laya/common.py`), not the `rl_common.py` shipped in the model
+//! repository, which has diverged from it.
+//!
+//! - [`config`] — the two config files a checkpoint ships.
+//! - [`backbone`] — ModernBERT, vendored so it runs outside f32.
+//! - [`head`] — type embedding, the head's transformer layers, the
+//!   marker scorer and the act head.
+//! - [`model`] — the composition, batch in, [`DecisionRow`]s out.
+
+pub mod backbone;
+pub mod config;
+pub mod head;
+pub mod model;
+
+pub use config::CheckpointConfig;
+pub use model::LayaModel;
+
+/// Question type, as the checkpoint's type embedding indexes it.
+pub const QTYPE_CHOICE: u8 = 0;
+pub const QTYPE_SCORE: u8 = 1;
+pub const QTYPE_NOUL: u8 = 2;
+
+/// One assembled question: its token ids, where each option's
+/// `[MASK]` marker sits, and its type.
+#[derive(Debug, Clone, PartialEq)]
+pub struct DecisionSeq {
+    pub input_ids: Vec<u32>,
+    pub markers: Vec<usize>,
+    /// [`QTYPE_CHOICE`], [`QTYPE_SCORE`] or [`QTYPE_NOUL`].
+    pub qtype: u8,
+}
+
+/// Sequences scored in one forward. One request's questions normally
+/// travel together, so the state they share is encoded in one pass.
+#[derive(Debug, Clone, Default, PartialEq)]
+pub struct DecisionBatch {
+    pub seqs: Vec<DecisionSeq>,
+}
+
+/// One sequence's result, CPU-side.
+#[derive(Debug, Clone, PartialEq)]
+pub struct DecisionRow {
+    /// One uncalibrated score per marker, in marker order.
+    pub logits: Vec<f32>,
+    /// Softmax over the act head's outputs. Index 0 is what the
+    /// reference reports as `action.act_probability`.
+    pub act: Vec<f32>,
+}
+
+#[cfg(test)]
+mod tests;

crates/neuron/src/harness/arch/laya/model.rs

@@ -0,0 +1,260 @@
+//! A whole Laya checkpoint: encoder + decision head, batch in, raw
+//! scores out.
+
+use anyhow::{Context, Result, ensure};
+use candle_core::{DType, Device, Tensor};
+use candle_nn::VarBuilder;
+use std::path::Path;
+
+use super::backbone::{Masks, ModernBert};
+use super::config::CheckpointConfig;
+use super::head::DecisionHead;
+use super::{DecisionBatch, DecisionRow};
+
+pub struct LayaModel {
+    encoder: ModernBert,
+    head: DecisionHead,
+    config: CheckpointConfig,
+    device: Device,
+    /// Precision of the linear layers (the load dtype).
+    dtype: DType,
+}
+
+/// A validated batch, flattened: what both forward paths start from.
+struct Flat {
+    lens: Vec<usize>,
+    qtypes: Vec<u32>,
+    /// Each row's marker positions, padded to `k_max` with 0.
+    marker_idx: Vec<u32>,
+    k_max: usize,
+    seq_len: usize,
+}
+
+impl LayaModel {
+    /// Load one checkpoint directory (the repo root, or one of its
+    /// subfolders): `rl_agent_config.json`, `encoder/config.json` and
+    /// `model.safetensors`.
+    pub fn load(dir: &Path, device: &Device, dtype: DType) -> Result<Self> {
+        let config = CheckpointConfig::from_dir(dir)?;
+        let weights = dir.join("model.safetensors");
+        // SAFETY: the file is memory-mapped read-only for the lifetime
+        // of the VarBuilder; nothing writes to a checkpoint in place.
+        let vb = unsafe { VarBuilder::from_mmaped_safetensors(&[&weights], dtype, device) }
+            .with_context(|| format!("mmap {}", weights.display()))?;
+        Self::from_var_builder(vb, config, device)
+    }
+
+    pub fn from_var_builder(
+        vb: VarBuilder,
+        config: CheckpointConfig,
+        device: &Device,
+    ) -> Result<Self> {
+        let dtype = vb.dtype();
+        let encoder = ModernBert::load(vb.pp("encoder"), &config.encoder)?;
+        let head = DecisionHead::load(
+            vb.clone(),
+            config.encoder.hidden_size,
+            config.laya.head_layers,
+            config.laya.n_act(),
+        )?;
+        Ok(Self {
+            encoder,
+            head,
+            config,
+            device: device.clone(),
+            dtype,
+        })
+    }
+
+    pub fn config(&self) -> &CheckpointConfig {
+        &self.config
+    }
+
+    /// Precision the linear layers were loaded in.
+    pub fn dtype(&self) -> DType {
+        self.dtype
+    }
+
+    /// One forward over every sequence in `batch`.
+    ///
+    /// On a CUDA device in reduced precision, in a build with the
+    /// `flash-attn` feature, the batch is packed without padding and
+    /// attention runs in flash varlen kernels. Otherwise rows are
+    /// right-padded to the longest, as the reference collates them, and
+    /// attention is eager; padding changes nothing about an unpadded
+    /// row's result.
+    pub fn forward(&self, batch: &DecisionBatch) -> Result<Vec<DecisionRow>> {
+        let flat = self.validate(batch)?;
+        #[cfg(feature = "flash-attn")]
+        if self.device.is_cuda() && self.dtype != DType::F32 && flash_enabled() {
+            return self.forward_packed(batch, &flat);
+        }
+        self.forward_padded(batch, &flat)
+    }
+
+    fn validate(&self, batch: &DecisionBatch) -> Result<Flat> {
+        let seqs = &batch.seqs;
+        ensure!(!seqs.is_empty(), "empty decision batch");
+        let max_pos = self.config.encoder.max_position_embeddings;
+        let seq_len = seqs.iter().map(|s| s.input_ids.len()).max().unwrap_or(0);
+        ensure!(seq_len > 0, "decision batch has an empty sequence");
+        ensure!(
+            seq_len <= max_pos,
+            "sequence of {seq_len} tokens exceeds max_position_embeddings {max_pos}"
+        );
+        let vocab = self.config.encoder.vocab_size as u32;
+        let k_max = seqs.iter().map(|s| s.markers.len()).max().unwrap_or(0);
+        ensure!(
+            k_max > 0,
+            "decision batch has a sequence with no option markers"
+        );
+        let mut lens = Vec::with_capacity(seqs.len());
+        let mut qtypes = Vec::with_capacity(seqs.len());
+        let mut marker_idx = Vec::with_capacity(seqs.len() * k_max);
+        for s in seqs {
+            let n = s.input_ids.len();
+            ensure!(n > 0, "decision batch has an empty sequence");
+            ensure!(!s.markers.is_empty(), "sequence has no option markers");
+            ensure!(s.qtype < 3, "question type {} out of range", s.qtype);
+            if let Some(&bad) = s.input_ids.iter().find(|&&t| t >= vocab) {
+                anyhow::bail!("token id {bad} outside vocabulary of {vocab}");
+            }
+            if let Some(&bad) = s.markers.iter().find(|&&m| m >= n) {
+                anyhow::bail!("marker position {bad} outside a {n}-token sequence");
+            }
+            lens.push(n);
+            qtypes.push(s.qtype as u32);
+            marker_idx.extend(s.markers.iter().map(|&m| m as u32));
+            // Absent slots gather position 0; their scores are discarded.
+            marker_idx.extend(std::iter::repeat_n(0u32, k_max - s.markers.len()));
+        }
+        Ok(Flat {
+            lens,
+            qtypes,
+            marker_idx,
+            k_max,
+            seq_len,
+        })
+    }
+
+    /// The eager path: rows right-padded to `(b, seq_len)`.
+    fn forward_padded(&self, batch: &DecisionBatch, flat: &Flat) -> Result<Vec<DecisionRow>> {
+        let b = batch.seqs.len();
+        let seq_len = flat.seq_len;
+        let pad = self.config.encoder.pad_token_id;
+        let mut ids = Vec::with_capacity(b * seq_len);
+        for s in &batch.seqs {
+            ids.extend_from_slice(&s.input_ids);
+            ids.extend(std::iter::repeat_n(pad, seq_len - s.input_ids.len()));
+        }
+        let ids = Tensor::from_vec(ids, (b, seq_len), &self.device)?;
+        let masks = Masks::new(
+            &flat.lens,
+            seq_len,
+            self.encoder.half_window(),
+            &self.device,
+        )?;
+        let qtype = Tensor::from_vec(flat.qtypes.clone(), b, &self.device)?;
+
+        let h = self.encoder.forward(&ids, &masks)?;
+        let h = self.head.contextualise(&h, &qtype, masks.global.as_ref())?;
+        let d = self.config.encoder.hidden_size;
+        let idx = Tensor::from_vec(flat.marker_idx.clone(), (b, flat.k_max, 1), &self.device)?
+            .broadcast_as((b, flat.k_max, d))?
+            .contiguous()?;
+        let marker_h = h.gather(&idx, 1)?;
+        let pooled = h.narrow(1, 0, 1)?.squeeze(1)?;
+        self.finish(batch, &marker_h, &pooled)
+    }
+
+    /// The flash path: tokens packed without padding, `(total, d)`.
+    #[cfg(feature = "flash-attn")]
+    fn forward_packed(&self, batch: &DecisionBatch, flat: &Flat) -> Result<Vec<DecisionRow>> {
+        use super::backbone::Packed;
+        let b = batch.seqs.len();
+        let packed = Packed::new(&flat.lens, &self.device)?;
+        let total: usize = flat.lens.iter().sum();
+        let mut ids = Vec::with_capacity(total);
+        let mut qtype_tok = Vec::with_capacity(total);
+        for (s, &q) in batch.seqs.iter().zip(&flat.qtypes) {
+            ids.extend_from_slice(&s.input_ids);
+            qtype_tok.extend(std::iter::repeat_n(q, s.input_ids.len()));
+        }
+        let ids = Tensor::from_vec(ids, total, &self.device)?;
+        let qtype_tok = Tensor::from_vec(qtype_tok, total, &self.device)?;
+
+        let h = self.encoder.forward_packed(&ids, &packed)?;
+        let h = self.head.contextualise_packed(&h, &qtype_tok, &packed)?;
+        let d = self.config.encoder.hidden_size;
+        // Marker positions, and each row's first token, in packed order.
+        let global: Vec<u32> = flat
+            .marker_idx
+            .chunks(flat.k_max)
+            .zip(&packed.offsets)
+            .flat_map(|(ms, &off)| ms.iter().map(move |&m| off as u32 + m))
+            .collect();
+        let global = Tensor::from_vec(global, b * flat.k_max, &self.device)?;
+        let marker_h = h.index_select(&global, 0)?.reshape((b, flat.k_max, d))?;
+        let firsts: Vec<u32> = packed.offsets.iter().map(|&o| o as u32).collect();
+        let pooled = h.index_select(&Tensor::from_vec(firsts, b, &self.device)?, 0)?;
+        self.finish(batch, &marker_h, &pooled)
+    }
+
+    /// Score the gathered marker states `(b, k_max, d)` and run the act
+    /// head over the pooled first positions `(b, d)`.
+    fn finish(
+        &self,
+        batch: &DecisionBatch,
+        marker_h: &Tensor,
+        pooled: &Tensor,
+    ) -> Result<Vec<DecisionRow>> {
+        let b = batch.seqs.len();
+        let logits: Vec<Vec<f32>> = self.head.score(marker_h)?.to_vec2()?;
+        let mut feats = Vec::with_capacity(b * 4);
+        let mut row_logits = Vec::with_capacity(b);
+        for (s, full) in batch.seqs.iter().zip(logits) {
+            let row: Vec<f32> = full[..s.markers.len()].to_vec();
+            feats.extend_from_slice(&act_features(&row));
+            row_logits.push(row);
+        }
+        let feats = Tensor::from_vec(feats, (b, 4), &self.device)?;
+        let act_logits = self.head.act(pooled, &feats)?;
+        let act = candle_nn::ops::softmax_last_dim(&act_logits)?.to_vec2::<f32>()?;
+        Ok(row_logits
+            .into_iter()
+            .zip(act)
+            .map(|(logits, act)| DecisionRow { logits, act })
+            .collect())
+    }
+}
+
+/// Runtime kill switch for the flash path: `NEURON_FLASH_ATTN=0` (or
+/// `false`) forces eager attention without a rebuild — the same switch
+/// the text architectures read. Read once.
+#[cfg(feature = "flash-attn")]
+fn flash_enabled() -> bool {
+    static ENABLED: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
+    *ENABLED
+        .get_or_init(|| !std::env::var("NEURON_FLASH_ATTN").is_ok_and(|v| v == "0" || v == "false"))
+}
+
+/// The act head's four inputs, from one row's uncalibrated logits:
+/// top-1 probability, top-1 minus top-2, entropy normalised by
+/// `ln(max(k, 2))`, and `k / 255`.
+///
+/// The reference computes these over the batch's padded width, with
+/// absent option slots filled with a logit of -1e4. Those slots carry
+/// exactly zero probability in f32, so computing per row over only the
+/// real options gives the same result.
+pub fn act_features(logits: &[f32]) -> [f32; 4] {
+    let max = logits.iter().copied().fold(f32::NEG_INFINITY, f32::max);
+    let exps: Vec<f32> = logits.iter().map(|&z| (z - max).exp()).collect();
+    let sum: f32 = exps.iter().sum();
+    let mut p: Vec<f32> = exps.iter().map(|e| e / sum).collect();
+    let k = (logits.len().max(2)) as f32;
+    let ent = -p.iter().map(|&x| x * x.max(1e-9).ln()).sum::<f32>() / k.ln();
+    p.sort_by(|a, b| b.total_cmp(a));
+    let top1 = p[0];
+    let top2 = p.get(1).copied().unwrap_or(0.0);
+    [top1, top1 - top2, ent, k / 255.0]
+}

crates/neuron/src/harness/arch/laya/tests.rs

@@ -0,0 +1,450 @@
+//! Parity against the upstream reference.
+//!
+//! `testdata/laya/tiny/` is a toy-sized checkpoint built and scored by
+//! the SDK's own `DecisionModel` (`script/generate_laya_fixture.py`);
+//! it runs in CI. `testdata/laya/reference.json` holds the released
+//! checkpoints' outputs over a real corpus (`script/laya-reference.py`);
+//! scoring it needs the weights, so that test is ignored unless
+//! `LAYA_REFERENCE_DIR` names a snapshot of `convaiinnovations/laya`.
+
+use super::*;
+use candle_core::{DType, Device};
+use serde::Deserialize;
+use std::path::{Path, PathBuf};
+
+fn testdata() -> PathBuf {
+    Path::new(env!("CARGO_MANIFEST_DIR")).join("src/harness/testdata/laya")
+}
+
+#[derive(Deserialize)]
+struct TinyExpected {
+    seqs: Vec<TinySeq>,
+}
+
+#[derive(Deserialize)]
+struct TinySeq {
+    input_ids: Vec<u32>,
+    markers: Vec<usize>,
+    qtype: u8,
+    logits: Vec<f32>,
+    act: Vec<f32>,
+}
+
+fn load_tiny() -> (LayaModel, TinyExpected) {
+    let dir = testdata().join("tiny");
+    let model = LayaModel::load(&dir, &Device::Cpu, DType::F32).expect("load tiny checkpoint");
+    let expected: TinyExpected = serde_json::from_str(
+        &std::fs::read_to_string(dir.join("expected.json")).expect("read expected.json"),
+    )
+    .expect("parse expected.json");
+    (model, expected)
+}
+
+fn batch_of(seqs: &[TinySeq]) -> DecisionBatch {
+    DecisionBatch {
+        seqs: seqs
+            .iter()
+            .map(|s| DecisionSeq {
+                input_ids: s.input_ids.clone(),
+                markers: s.markers.clone(),
+                qtype: s.qtype,
+            })
+            .collect(),
+    }
+}
+
+/// Device and precision for the weight-backed tests: `LAYA_DEVICE`
+/// (`cpu`, or `cuda:N` in a build with the `cuda` feature) and
+/// `LAYA_DTYPE` (`f32`, `bf16` or `f16`). Defaults: CPU, f32.
+fn test_device() -> (Device, DType) {
+    let device = match std::env::var("LAYA_DEVICE").as_deref() {
+        Ok("cpu") | Err(_) => Device::Cpu,
+        Ok(s) => {
+            let ordinal = s
+                .strip_prefix("cuda:")
+                .and_then(|n| n.parse().ok())
+                .unwrap_or_else(|| panic!("LAYA_DEVICE={s}: expected cpu or cuda:N"));
+            Device::new_cuda(ordinal).unwrap_or_else(|e| panic!("LAYA_DEVICE={s}: {e}"))
+        }
+    };
+    let dtype = match std::env::var("LAYA_DTYPE").as_deref() {
+        Ok("f32") | Err(_) => DType::F32,
+        Ok("bf16") => DType::BF16,
+        Ok("f16") => DType::F16,
+        Ok(s) => panic!("LAYA_DTYPE={s}: expected f32, bf16 or f16"),
+    };
+    (device, dtype)
+}
+
+/// Logit tolerance against the f32 reference for a given precision.
+/// Reduced precision is judged mainly on argmax agreement; the bound
+/// only catches a broken kernel, not rounding.
+fn logit_tolerance(dtype: DType) -> f32 {
+    if dtype == DType::F32 { 1e-3 } else { 1.0 }
+}
+
+fn max_abs_diff(a: &[f32], b: &[f32]) -> f32 {
+    assert_eq!(a.len(), b.len(), "length mismatch");
+    a.iter()
+        .zip(b)
+        .map(|(x, y)| (x - y).abs())
+        .fold(0.0, f32::max)
+}
+
+/// The whole forward — encoder with both attention kinds and rotary
+/// bases, padding, type embedding, head layers, marker gather, scorer
+/// and act head — against the reference, in f32.
+#[test]
+fn tiny_checkpoint_matches_reference() {
+    let (model, expected) = load_tiny();
+    let rows = model.forward(&batch_of(&expected.seqs)).expect("forward");
+    assert_eq!(rows.len(), expected.seqs.len());
+    for (i, (got, want)) in rows.iter().zip(&expected.seqs).enumerate() {
+        let dl = max_abs_diff(&got.logits, &want.logits);
+        let da = max_abs_diff(&got.act, &want.act);
+        assert!(
+            dl < 1e-5 && da < 1e-5,
+            "row {i}: logit diff {dl:e}, act diff {da:e}\n got {:?} / {:?}\nwant {:?} / {:?}",
+            got.logits,
+            got.act,
+            want.logits,
+            want.act
+        );
+    }
+}
+
+/// A row's result must not depend on what it was batched with: padded
+/// keys are masked out of every attention, including the head's.
+#[test]
+fn padding_does_not_change_a_row() {
+    let (model, expected) = load_tiny();
+    let batched = model.forward(&batch_of(&expected.seqs)).expect("batched");
+    for (i, seq) in expected.seqs.iter().enumerate() {
+        let alone = model
+            .forward(&batch_of(std::slice::from_ref(seq)))
+            .expect("single");
+        let dl = max_abs_diff(&alone[0].logits, &batched[i].logits);
+        let da = max_abs_diff(&alone[0].act, &batched[i].act);
+        assert!(dl < 1e-5 && da < 1e-5, "row {i}: {dl:e} / {da:e}");
+    }
+}
+
+#[test]
+fn malformed_batches_are_refused() {
+    let (model, _) = load_tiny();
+    let seq = |ids: Vec<u32>, markers: Vec<usize>, qtype| DecisionBatch {
+        seqs: vec![DecisionSeq {
+            input_ids: ids,
+            markers,
+            qtype,
+        }],
+    };
+    assert!(model.forward(&DecisionBatch::default()).is_err());
+    assert!(model.forward(&seq(vec![1, 2, 3], vec![], 0)).is_err());
+    assert!(model.forward(&seq(vec![1, 2, 3], vec![3], 0)).is_err());
+    assert!(model.forward(&seq(vec![1, 2, 3], vec![1], 3)).is_err());
+    assert!(model.forward(&seq(vec![1, 500, 3], vec![1], 0)).is_err());
+}
+
+#[test]
+fn act_features_match_the_reference_definition() {
+    // Two equal logits: top1 = top2 = 0.5, entropy exactly 1.
+    let f = model::act_features(&[0.3, 0.3]);
+    assert!((f[0] - 0.5).abs() < 1e-6 && f[1].abs() < 1e-6 && (f[2] - 1.0).abs() < 1e-6);
+    assert!((f[3] - 2.0 / 255.0).abs() < 1e-7);
+    // One option: probability 1, no second, k clamps to 2.
+    let f = model::act_features(&[4.0]);
+    assert_eq!(f[0], 1.0);
+    assert_eq!(f[1], 1.0);
+    assert!(f[2].abs() < 1e-6);
+    assert!((f[3] - 2.0 / 255.0).abs() < 1e-7);
+}
+
+// ---------------------------------------------------------------- released checkpoints
+
+#[derive(Deserialize)]
+struct Reference {
+    cases: Vec<RefCase>,
+}
+
+#[derive(Deserialize)]
+struct RefCase {
+    name: String,
+    items: Vec<RefItem>,
+    response: serde_json::Value,
+}
+
+#[derive(Deserialize)]
+struct RefItem {
+    input_ids: Vec<u32>,
+    markers: Vec<usize>,
+    qtype: u8,
+    logits: Vec<f32>,
+    act: Vec<f32>,
+}
+
+/// Every case in `reference.json`, scored by the checkpoint the
+/// reference routed it to, on the device and precision
+/// [`test_device`] selects (f32 on the CPU by default).
+///
+/// `LAYA_REFERENCE_DIR` is a snapshot of `convaiinnovations/laya` at the
+/// revision recorded in the fixture header: the root holds the English
+/// checkpoint, `multilingual/` and `typed-decisions/` the other two.
+#[test]
+#[ignore = "needs the released weights: set LAYA_REFERENCE_DIR"]
+fn released_checkpoints_match_reference() {
+    let Ok(root) = std::env::var("LAYA_REFERENCE_DIR") else {
+        panic!("LAYA_REFERENCE_DIR is not set");
+    };
+    let root = PathBuf::from(root);
+    let (device, dtype) = test_device();
+    let tolerance = logit_tolerance(dtype);
+    let reference: Reference = serde_json::from_str(
+        &std::fs::read_to_string(testdata().join("reference.json")).expect("read reference.json"),
+    )
+    .expect("parse reference.json");
+
+    // The reference's own GPU run (bf16 autocast), for judging reduced
+    // precision against the drift upstream itself serves with.
+    let torch_bf16: Option<Reference> =
+        std::fs::read_to_string(testdata().join("reference_cuda_bf16.json"))
+            .ok()
+            .map(|t| serde_json::from_str(&t).expect("parse reference_cuda_bf16.json"));
+    // Per checkpoint: summed centred drift of ours and of the reference's
+    // bf16 run, their maxima, and the row count.
+    let mut drift: std::collections::BTreeMap<String, (f32, f32, f32, f32, usize)> =
+        Default::default();
+
+    let mut models = std::collections::BTreeMap::new();
+    let mut stats: std::collections::BTreeMap<String, (usize, f32, f32, usize)> =
+        Default::default();
+    // Forward wall time and tokens encoded, per checkpoint: a first
+    // CPU number, not a benchmark (debug builds are far slower).
+    let mut timing: std::collections::BTreeMap<String, (f64, usize)> = Default::default();
+    for case in &reference.cases {
+        let ckpt = case.response["routing"]["model"]
+            .as_str()
+            .expect("routing.model")
+            .to_string();
+        let model = models.entry(ckpt.clone()).or_insert_with(|| {
+            let dir = match ckpt.as_str() {
+                "english" => root.clone(),
+                other => root.join(other),
+            };
+            LayaModel::load(&dir, &device, dtype).unwrap_or_else(|e| panic!("load {ckpt}: {e:#}"))
+        });
+        let batch = DecisionBatch {
+            seqs: case
+                .items
+                .iter()
+                .map(|it| DecisionSeq {
+                    input_ids: it.input_ids.clone(),
+                    markers: it.markers.clone(),
+                    qtype: it.qtype,
+                })
+                .collect(),
+        };
+        let started = std::time::Instant::now();
+        let rows = model
+            .forward(&batch)
+            .unwrap_or_else(|e| panic!("{}: {e:#}", case.name));
+        let t = timing.entry(ckpt.clone()).or_default();
+        t.0 += started.elapsed().as_secs_f64() * 1e3;
+        t.1 += batch.seqs.iter().map(|s| s.input_ids.len()).sum::<usize>();
+        let torch_case = torch_bf16
+            .as_ref()
+            .and_then(|r| r.cases.iter().find(|c| c.name == case.name));
+        for (i, (got, want)) in rows.iter().zip(&case.items).enumerate() {
+            let d = drift.entry(ckpt.clone()).or_default();
+            let ours = centred_diff(&got.logits, &want.logits);
+            d.0 += ours;
+            d.2 = d.2.max(ours);
+            if let Some(t) = torch_case.and_then(|c| c.items.get(i)) {
+                let theirs = centred_diff(&t.logits, &want.logits);
+                d.1 += theirs;
+                d.3 = d.3.max(theirs);
+            }
+            d.4 += 1;
+        }
+        let entry = stats.entry(ckpt.clone()).or_default();
+        for (got, want) in rows.iter().zip(&case.items) {
+            let dl = max_abs_diff(&got.logits, &want.logits);
+            let da = max_abs_diff(&got.act, &want.act);
+            let argmax = |v: &[f32]| {
+                v.iter()
+                    .enumerate()
+                    .max_by(|a, b| a.1.total_cmp(b.1))
+                    .map(|(i, _)| i)
+            };
+            entry.0 += 1;
+            entry.1 = entry.1.max(dl);
+            entry.2 = entry.2.max(da);
+            if argmax(&got.logits) == argmax(&want.logits) {
+                entry.3 += 1;
+            }
+            if dtype != DType::F32 && std::env::var_os("LAYA_VERBOSE").is_some() {
+                eprintln!("  {:<40} {ckpt:<16} |Δlogit| {dl:.3}", case.name);
+            }
+            assert!(
+                dl < tolerance,
+                "{} ({ckpt}): logit diff {dl:e}\n got {:?}\nwant {:?}",
+                case.name,
+                got.logits,
+                want.logits
+            );
+            assert!(da < tolerance, "{} ({ckpt}): act diff {da:e}", case.name);
+        }
+    }
+    for (ckpt, (n, dl, da, agree)) in &stats {
+        let (ms, tokens) = timing[ckpt];
+        eprintln!(
+            "[{device:?} {dtype:?}] {ckpt}: {n} questions, max |Δlogit| {dl:.2e}, max |Δact| {da:.2e}, argmax {agree}/{n}; \
+             {tokens} tokens in {ms:.0} ms ({:.2} ms/token)",
+            ms / tokens as f64
+        );
+        assert_eq!(agree, n, "{ckpt}: argmax disagreement");
+    }
+    // Reduced precision: centred drift (softmax ignores a constant shift)
+    // against f32, next to the reference's own bf16 drift. Per-case drift
+    // is chaotic, so the gate is the mean: no worse than twice upstream's,
+    // plus 0.01 — about one bf16 ulp at these logits' magnitude — so a
+    // checkpoint with a single fixture case whose reference drift happens
+    // to round to nothing is not held to zero.
+    for (ckpt, (ours, theirs, ours_max, theirs_max, n)) in &drift {
+        let (ours, theirs) = (ours / *n as f32, theirs / *n as f32);
+        eprintln!(
+            "[{device:?} {dtype:?}] {ckpt}: centred drift mean {ours:.4} (reference bf16 {theirs:.4}), \
+             max {ours_max:.3} (reference bf16 {theirs_max:.3})"
+        );
+        if dtype != DType::F32 && torch_bf16.is_some() {
+            assert!(
+                ours <= 2.0 * theirs + 0.01,
+                "{ckpt}: mean drift {ours:.4} exceeds twice the reference's own bf16 drift {theirs:.4}"
+            );
+        }
+    }
+}
+
+/// Largest difference between two logit rows after removing each row's
+/// mean — the part of a difference a softmax can see.
+fn centred_diff(a: &[f32], b: &[f32]) -> f32 {
+    let mean = |v: &[f32]| v.iter().sum::<f32>() / v.len() as f32;
+    let (ma, mb) = (mean(a), mean(b));
+    a.iter()
+        .zip(b)
+        .map(|(x, y)| ((x - ma) - (y - mb)).abs())
+        .fold(0.0, f32::max)
+}
+
+/// Forward latency for the four workloads #337 benchmarks against the
+/// PyTorch reference, plus batch throughput, on the English checkpoint.
+///
+/// Sequences are real ones from `reference.json`: `choice_basic` is the
+/// short state (41 tokens), `multi_question` five questions over it, and
+/// `state_right_truncated` a full 512-token sequence (repeated for the
+/// five-question and throughput batches). The timed span is the whole
+/// `forward` call — host-to-device copy, forward, and the CPU readback
+/// that ends it — so it is directly comparable to the reference's
+/// end-to-end `predict()` minus tokenisation.
+///
+/// Run with `cargo test --release -p neuron [--features cuda] --
+/// --ignored --nocapture laya_forward_timing`, with `LAYA_REFERENCE_DIR`
+/// and optionally `LAYA_DEVICE` / `LAYA_DTYPE` set.
+#[test]
+#[ignore = "benchmark: needs the released weights (LAYA_REFERENCE_DIR)"]
+fn laya_forward_timing() {
+    let root = PathBuf::from(std::env::var("LAYA_REFERENCE_DIR").expect("LAYA_REFERENCE_DIR"));
+    let (device, dtype) = test_device();
+    let reference: Reference = serde_json::from_str(
+        &std::fs::read_to_string(testdata().join("reference.json")).expect("read reference.json"),
+    )
+    .expect("parse reference.json");
+    let seqs_of = |name: &str| -> Vec<DecisionSeq> {
+        reference
+            .cases
+            .iter()
+            .find(|c| c.name == name)
+            .unwrap_or_else(|| panic!("no case {name}"))
+            .items
+            .iter()
+            .map(|it| DecisionSeq {
+                input_ids: it.input_ids.clone(),
+                markers: it.markers.clone(),
+                qtype: it.qtype,
+            })
+            .collect()
+    };
+    let long = seqs_of("state_right_truncated").remove(0);
+    let workloads: Vec<(String, DecisionBatch)> = vec![
+        (
+            "short x1q".into(),
+            DecisionBatch {
+                seqs: seqs_of("choice_basic"),
+            },
+        ),
+        (
+            "short x5q".into(),
+            DecisionBatch {
+                seqs: seqs_of("multi_question"),
+            },
+        ),
+        (
+            "512tok x1q".into(),
+            DecisionBatch {
+                seqs: vec![long.clone()],
+            },
+        ),
+        (
+            "512tok x5q".into(),
+            DecisionBatch {
+                seqs: vec![long.clone(); 5],
+            },
+        ),
+        (
+            "512tok x16".into(),
+            DecisionBatch {
+                seqs: vec![long.clone(); 16],
+            },
+        ),
+        (
+            "512tok x64".into(),
+            DecisionBatch {
+                seqs: vec![long.clone(); 64],
+            },
+        ),
+    ];
+
+    let model = LayaModel::load(&root, &device, dtype).expect("load English checkpoint");
+    // `LAYA_TIMING_ONLY` keeps the workloads whose name contains it, and
+    // `LAYA_TIMING_ITERS` sets the timed iterations — for profiling one
+    // workload under a tracer.
+    let only = std::env::var("LAYA_TIMING_ONLY").ok();
+    let iters: usize = std::env::var("LAYA_TIMING_ITERS")
+        .ok()
+        .and_then(|v| v.parse().ok())
+        .unwrap_or(20);
+    for (name, batch) in &workloads {
+        if only.as_deref().is_some_and(|o| !name.contains(o)) {
+            continue;
+        }
+        for _ in 0..3 {
+            model.forward(batch).expect("warm-up forward");
+        }
+        let mut ms: Vec<f64> = (0..iters)
+            .map(|_| {
+                let started = std::time::Instant::now();
+                model.forward(batch).expect("forward");
+                started.elapsed().as_secs_f64() * 1e3
+            })
+            .collect();
+        ms.sort_by(f64::total_cmp);
+        let median = ms[ms.len() / 2];
+        let tokens: usize = batch.seqs.iter().map(|s| s.input_ids.len()).sum();
+        eprintln!(
+            "[{device:?} {dtype:?}] {name:>11}: median {median:7.1} ms, min {:7.1} ms, \
+             {tokens} tokens, {:.0} seq/s",
+            ms[0],
+            batch.seqs.len() as f64 / (median / 1e3)
+        );
+    }
+}

crates/neuron/src/harness/arch/mod.rs

@@ -20,6 +20,8 @@
 //! TP-aware analogues live in `harness/tp/tp_<family>.rs` and follow
 //! the pattern set by `tp_qwen3.rs`.
 
+pub mod laya;
+pub mod phase_probe;
 pub mod qwen3_5;
 pub mod qwen4_exp;
 pub mod snapshot;

crates/neuron/src/harness/arch/phase_probe.rs

@@ -0,0 +1,167 @@
+//! Opt-in phase timing inside a model's forward pass.
+//!
+//! The request-level split (`PhaseTiming` on the wire) says whether a
+//! decode step's time went into the forward, the sampler or the emit.
+//! When the answer is "the forward", this says *where* in the forward —
+//! embedding, a gather, the layer stack, the head.
+//!
+//! # Why this needs a device synchronise, and why it is off by default
+//!
+//! candle's CUDA ops are asynchronous on the stream. An `Instant` taken
+//! around GPU work therefore measures how fast kernels were *queued*,
+//! not how long they ran, and whichever phase happens to contain the
+//! next real synchronisation point silently absorbs everyone else's
+//! cost. The numbers look plausible and attribute the time to the wrong
+//! place — a worse failure than no numbers at all.
+//!
+//! So each mark synchronises the device first. That is only affordable
+//! when the phases being measured are much larger than a sync: at a few
+//! tens of microseconds per sync, a handful of marks is noise against a
+//! step measured in hundreds of milliseconds, and material against one
+//! measured in single milliseconds. Since that ratio is a property of
+//! how slow the model currently is rather than something safe to assume,
+//! the probe stays off unless `NEURON_FORWARD_PHASES` is set.
+
+use candle_core::Device;
+use std::sync::OnceLock;
+use std::time::Instant;
+
+fn enabled() -> bool {
+    static ON: OnceLock<bool> = OnceLock::new();
+    *ON.get_or_init(|| std::env::var("NEURON_FORWARD_PHASES").is_ok())
+}
+
+/// Accumulates named phases within one forward pass.
+///
+/// Construct at the top of the forward, call [`mark`] at each boundary,
+/// and [`emit`] at the end. Disabled, every method is a predictable
+/// branch and nothing is allocated.
+///
+/// [`mark`]: PhaseProbe::mark
+/// [`emit`]: PhaseProbe::emit
+pub struct PhaseProbe {
+    on: bool,
+    device: Device,
+    last: Instant,
+    marks: Vec<(&'static str, f64)>,
+}
+
+impl PhaseProbe {
+    pub fn new(device: &Device) -> Self {
+        Self::with_enabled(device, enabled())
+    }
+
+    /// Constructor with the gate supplied explicitly.
+    ///
+    /// [`new`] reads the environment through a `OnceLock`, which a test
+    /// cannot toggle — so without this the enabled path would be
+    /// unreachable from a test, and a probe that only ever runs
+    /// disabled is not evidence that it records anything.
+    ///
+    /// [`new`]: PhaseProbe::new
+    pub fn with_enabled(device: &Device, on: bool) -> Self {
+        // Start from a synchronised device, or the first phase inherits
+        // whatever was still in flight when the forward was entered.
+        if on {
+            let _ = device.synchronize();
+        }
+        Self {
+            on,
+            device: device.clone(),
+            last: Instant::now(),
+            marks: Vec::new(),
+        }
+    }
+
+    /// Close the phase that ended here and open the next one.
+    pub fn mark(&mut self, name: &'static str) {
+        if !self.on {
+            return;
+        }
+        // A failed synchronise means the elapsed time below is not the
+        // phase's own. Record it as such rather than reporting a number
+        // that looks like a measurement.
+        let synced = self.device.synchronize().is_ok();
+        let ms = self.last.elapsed().as_secs_f64() * 1e3;
+        self.marks
+            .push((if synced { name } else { "UNSYNCED" }, ms));
+        self.last = Instant::now();
+    }
+
+    /// The phases recorded so far, in order.
+    #[cfg(test)]
+    pub fn marks(&self) -> &[(&'static str, f64)] {
+        &self.marks
+    }
+
+    /// Emit one line carrying every phase of this forward.
+    pub fn emit(self, rows: usize) {
+        if !self.on || self.marks.is_empty() {
+            return;
+        }
+        let total: f64 = self.marks.iter().map(|(_, ms)| ms).sum();
+        let detail = self
+            .marks
+            .iter()
+            .map(|(n, ms)| format!("{n}={ms:.3}"))
+            .collect::<Vec<_>>()
+            .join(" ");
+        tracing::info!(rows, total_ms = total, %detail, "forward phases");
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    /// A disabled probe must record nothing at all — not zero-valued
+    /// phases. Zeros in a log read as "this phase was free", which is
+    /// the same failure the wire-side `Option` exists to prevent.
+    #[test]
+    fn a_disabled_probe_records_nothing() {
+        let mut p = PhaseProbe::with_enabled(&Device::Cpu, false);
+        p.mark("embed");
+        p.mark("layers");
+        assert!(
+            p.marks().is_empty(),
+            "disabled probe recorded {:?}",
+            p.marks()
+        );
+    }
+
+    /// An enabled probe records one phase per mark, in call order, and
+    /// each phase measures only the span since the previous mark rather
+    /// than since the start.
+    #[test]
+    fn phases_are_recorded_in_order_and_do_not_accumulate() {
+        let mut p = PhaseProbe::with_enabled(&Device::Cpu, true);
+        p.mark("embed");
+        std::thread::sleep(std::time::Duration::from_millis(20));
+        p.mark("layers");
+        p.mark("head");
+
+        let names: Vec<_> = p.marks().iter().map(|(n, _)| *n).collect();
+        assert_eq!(names, ["embed", "layers", "head"]);
+
+        let ms = |i: usize| p.marks()[i].1;
+        // "layers" contains the sleep; the marks either side must not.
+        // A probe that timed from the start instead of from the previous
+        // mark would make "head" >= "layers", which is the mistake this
+        // pins.
+        assert!(
+            ms(1) >= 15.0,
+            "layers should carry the sleep, got {:?}",
+            p.marks()
+        );
+        assert!(
+            ms(2) < 10.0,
+            "head must not inherit it, got {:?}",
+            p.marks()
+        );
+        assert!(
+            ms(0) < 10.0,
+            "embed must not inherit it, got {:?}",
+            p.marks()
+        );
+    }
+}

crates/neuron/src/harness/arch/qwen3_5/decoder.rs

@@ -77,7 +77,25 @@ impl Qwen3_5DecoderLayer {
                     cfg.layer_types.len()
                 )
             })?;
+        Self::load_typed(cfg, rotary, layer_type, cfg.layer_uses_moe(layer_idx), vb)
+    }
 
+    /// Load a layer whose flavour is stated rather than looked up.
+    ///
+    /// The MTP head's layer (#96) sits at `mtp.layers.0`, outside the
+    /// `layer_types` list that indexes the target's own layers — asking
+    /// that list about it would answer for a target layer that happens
+    /// to share the index. The checkpoint says what it is: plain
+    /// `self_attn.{q,k,v,o}_proj` with `q_norm`/`k_norm`, and a dense
+    /// `mlp`, i.e. full attention without MoE, whatever the target
+    /// layer at index 0 happens to be.
+    pub fn load_typed(
+        cfg: &TextConfig,
+        rotary: Arc<RotaryEmbedding>,
+        layer_type: &str,
+        use_moe: bool,
+        vb: &ShardedVarBuilder,
+    ) -> Result<Self> {
         let attention = match layer_type {
             "full_attention" => {
                 AttentionKind::Full(Qwen3_5Attention::load(cfg, rotary, &vb.pp("self_attn"))?)
@@ -86,12 +104,12 @@ impl Qwen3_5DecoderLayer {
                 AttentionKind::Linear(GatedDeltaNet::load(cfg, &vb.pp("linear_attn"))?)
             }
             other => anyhow::bail!(
-                "unknown layer_type '{other}' for layer {layer_idx} (expected \
-                 'full_attention' or 'linear_attention')"
+                "unknown layer_type '{other}' (expected 'full_attention' or \
+                 'linear_attention')"
             ),
         };
 
-        let mlp = if cfg.layer_uses_moe(layer_idx) {
+        let mlp = if use_moe {
             MlpKind::Moe(Qwen3_5MoeBlock::load(cfg, &vb.pp("mlp"))?)
         } else {
             MlpKind::Dense(Qwen3_5MLP::load(cfg, &vb.pp("mlp"))?)

crates/neuron/src/harness/arch/qwen3_5/linear_attn.rs

@@ -1391,6 +1391,8 @@ mod tests {
         let (b, l) = (1, 3);
         let cfg = TextConfig {
             vocab_size: 100,
+            mtp_num_hidden_layers: 0,
+            mtp_use_dedicated_embeddings: false,
             hidden_size: 16,
             intermediate_size: 32,
             num_hidden_layers: 1,

crates/neuron/src/harness/arch/qwen3_5/mod.rs

@@ -77,6 +77,7 @@ pub mod full_attn;
 pub mod linear_attn;
 pub mod mlp;
 pub mod moe;
+pub mod mtp;
 pub mod rmsnorm;
 pub mod rope;
 pub mod vision;
@@ -225,6 +226,16 @@ pub struct TextConfig {
     /// pointwise after softmax+matmul.
     #[serde(default)]
     pub attn_output_gate: bool,
+    /// Multi-token-prediction head depth (#96). `1` on the checkpoints
+    /// that ship one (`Qwen3.8-27B`, `Qwen3.5-0.8B`); absent or `0`
+    /// means the weights carry no `mtp.*` tensors.
+    #[serde(default)]
+    pub mtp_num_hidden_layers: usize,
+    /// Whether the MTP head has its own embedding table. `false` on
+    /// every checkpoint seen so far, which is why the head costs only
+    /// its own layer and not a second vocabulary-sized matrix.
+    #[serde(default)]
+    pub mtp_use_dedicated_embeddings: bool,
 
     /// One entry per decoder layer; values are `"full_attention"` or
     /// `"linear_attention"`. Length must equal `num_hidden_layers`.
@@ -883,6 +894,53 @@ impl Qwen3_5ForCausalLM {
         hidden.i((.., l - 1.., ..))?.apply(&self.lm_head)
     }
 
+    /// Logits at **every** position: `(B, L, vocab)`.
+    ///
+    /// [`Self::forward`] slices the last position off before the LM
+    /// head, which is what a plain decode step wants. Speculative
+    /// verification (#96) wants the opposite: the target forwards the
+    /// drafted block in one pass and needs its own token at each
+    /// position to compare against the draft. The forward already
+    /// computes all of them — chunked prefill relies on it — so this
+    /// only declines to throw them away.
+    pub fn forward_multi(&mut self, input: &Tensor, offset: usize) -> candle_core::Result<Tensor> {
+        let hidden = self.base.forward(input, offset)?;
+        hidden.apply(&self.lm_head)
+    }
+
+    /// Like [`Self::forward`], but also returns the hidden state the
+    /// logits were produced from — `(B, L, hidden)`, post-final-norm,
+    /// every position.
+    ///
+    /// The MTP head (#96) consumes exactly this tensor: vLLM's
+    /// `Qwen3NextModel::forward` returns `self.norm(...)` and its
+    /// runner hands that to the draft head. Kept beside `forward` so
+    /// the two cannot drift about which hidden state that is.
+    pub fn forward_with_hidden(
+        &mut self,
+        input: &Tensor,
+        offset: usize,
+    ) -> candle_core::Result<(Tensor, Tensor)> {
+        let (_, l) = input.dims2()?;
+        let hidden = self.base.forward(input, offset)?;
+        let logits = hidden.i((.., l - 1.., ..))?.apply(&self.lm_head)?;
+        Ok((logits, hidden))
+    }
+
+    /// Apply the LM head to a hidden state. The MTP head has none of
+    /// its own (#96) — draft and target logits come from this same
+    /// projection, which is what makes their distributions comparable.
+    pub fn lm_head(&self, hidden: &Tensor) -> candle_core::Result<Tensor> {
+        hidden.apply(&self.lm_head)
+    }
+
+    /// The token embedding the MTP head fuses with the hidden state.
+    /// The head has none of its own (`mtp_use_dedicated_embeddings` is
+    /// false on every checkpoint seen), so it borrows the target's.
+    pub fn embed_tokens(&self, input: &Tensor) -> candle_core::Result<Tensor> {
+        self.base.embed_tokens.forward(input)
+    }
+
     /// Lockstep batched decode step (#98): `(B, 1)` input, per-row
     /// positions, padding mask from
     /// [`Qwen3_5Model::batch_decode_mask`]. Returns `(B, 1,

crates/neuron/src/harness/arch/qwen3_5/mtp.rs

@@ -0,0 +1,485 @@
+//! The multi-token-prediction head — `qwen3_5`'s shipped draft model.
+//!
+//! `Qwen/Qwen3.8-27B` and `Qwen/Qwen3.5-0.8B` both carry a complete
+//! extra decoder layer under `mtp.*` that nothing in this crate has
+//! ever loaded: 15 tensors of 1199 in the 27B (810 MB BF16), 15 of 488
+//! in the 0.8B (39 MB). It is the drafter speculative decoding needs
+//! (#96), already in the weights we download.
+//!
+//! ```text
+//! e = pre_fc_norm_embedding(embed(t))      [B, L, H]
+//! h = pre_fc_norm_hidden(h)                [B, L, H]  (the target's)
+//! x = fc(cat[e, h])                        [B, L, 2H] -> [B, L, H]
+//! x = layer(x)                             one full-attention block
+//! x = norm(x)                              -> the target's lm_head
+//! ```
+//!
+//! Three things here cannot be recovered from the tensor shapes, and
+//! all three were read from vLLM's `qwen3_next_mtp.py` rather than
+//! inferred — the same discipline, and for the same reason, as the
+//! sibling head in `arch/qwen4_exp/mtp.rs`:
+//!
+//! 1. **The concatenation is `[embedding, hidden]`**, not the reverse:
+//!    `hidden_states = torch.cat([inputs_embeds, hidden_states], -1)`.
+//!    `fc` is square in the sense that both halves are `hidden_size`
+//!    wide, so swapping them loads, runs, and produces finite logits
+//!    that are simply wrong. Pinned by value below.
+//! 2. **Each half is normalised before the concatenation**, by its own
+//!    norm — `pre_fc_norm_embedding` on the embedding,
+//!    `pre_fc_norm_hidden` on the hidden state — never one norm over
+//!    the joined 2H.
+//! 3. **The hidden state is the target's *post-final-norm* output**,
+//!    the same tensor its `lm_head` consumes. vLLM's
+//!    `Qwen3NextModel::forward` returns `self.norm(...)` and the runner
+//!    hands that to the head. Feeding the pre-norm residual instead is
+//!    the other plausible reading, and it is wrong for this
+//!    architecture (`qwen4_exp`'s head takes its pre-mixer stream —
+//!    they do not generalise to each other).
+//!
+//! **The head has no `lm_head` or `embed_tokens` of its own.** The
+//! checkpoint carries neither (`mtp_use_dedicated_embeddings: false`),
+//! so both come from the target — which is what makes the resident cost
+//! the 810 MB above rather than that plus a second embedding table.
+//!
+//! **The head keeps its own KV cache, and it is not the target's.** A
+//! head fed only the newest token attends over an empty past, so it has
+//! to walk the prompt alongside the target before it can draft. That
+//! puts MTP prefill inside the speculative loop's cost, not just decode
+//! — see #96.
+
+use anyhow::{Context, Result};
+use candle_core::{Module, Tensor};
+use candle_nn::Linear;
+use candle_nn::var_builder::ShardedVarBuilder;
+use std::sync::Arc;
+
+use super::TextConfig;
+use super::decoder::Qwen3_5DecoderLayer;
+use super::rmsnorm::Qwen3_5RmsNorm;
+use super::rope::RotaryEmbedding;
+
+pub struct MtpHead {
+    pre_fc_norm_embedding: Qwen3_5RmsNorm,
+    pre_fc_norm_hidden: Qwen3_5RmsNorm,
+    /// `Linear(2H -> H)` over `cat[embedding, hidden]`.
+    fc: Linear,
+    layer: Qwen3_5DecoderLayer,
+    /// Final norm before the *target's* lm_head.
+    norm: Qwen3_5RmsNorm,
+}
+
+impl MtpHead {
+    /// `vb` is the checkpoint root; the head hangs off `mtp.`.
+    pub fn load(
+        cfg: &TextConfig,
+        rotary: Arc<RotaryEmbedding>,
+        vb: &ShardedVarBuilder,
+    ) -> Result<Self> {
+        let mtp = vb.pp("mtp");
+        let h = cfg.hidden_size;
+        let fc_weight = mtp
+            .pp("fc")
+            .get((h, 2 * h), "weight")
+            .context("load 'mtp.fc.weight'")?;
+        Ok(Self {
+            pre_fc_norm_embedding: Qwen3_5RmsNorm::load(
+                &mtp.pp("pre_fc_norm_embedding"),
+                h,
+                cfg.rms_norm_eps,
+            )
+            .context("load mtp.pre_fc_norm_embedding")?,
+            pre_fc_norm_hidden: Qwen3_5RmsNorm::load(
+                &mtp.pp("pre_fc_norm_hidden"),
+                h,
+                cfg.rms_norm_eps,
+            )
+            .context("load mtp.pre_fc_norm_hidden")?,
+            fc: Linear::new(fc_weight, None),
+            // Stated, not looked up: the head's layer is outside
+            // `layer_types`. Full attention, dense MLP — see
+            // `Qwen3_5DecoderLayer::load_typed`.
+            layer: Qwen3_5DecoderLayer::load_typed(
+                cfg,
+                rotary,
+                "full_attention",
+                false,
+                &mtp.pp("layers").pp(0),
+            )
+            .context("load mtp.layers.0")?,
+            norm: Qwen3_5RmsNorm::load(&mtp.pp("norm"), h, cfg.rms_norm_eps)
+                .context("load mtp.norm")?,
+        })
+    }
+
+    /// Whether this checkpoint carries a head at all.
+    ///
+    /// `mtp_num_hidden_layers` is the config's own answer and is 1 on
+    /// the checkpoints that ship one. Absent or 0 means the weights
+    /// have no `mtp.*` tensors and [`Self::load`] would fail.
+    pub fn present_in(cfg: &TextConfig) -> bool {
+        cfg.mtp_num_hidden_layers > 0
+    }
+
+    /// One draft step.
+    ///
+    /// `token_embed` is `embed_tokens(t)` for the token the draft is
+    /// predicting *from*, `[B, L, H]`. `hidden` is the target's
+    /// post-final-norm hidden state at the same positions, `[B, L, H]`
+    /// — on a chained second step, this head's own previous output.
+    /// The returned tensor is `[B, L, H]`, for the *target's* lm_head.
+    pub fn forward(
+        &mut self,
+        token_embed: &Tensor,
+        hidden: &Tensor,
+        attn_mask: Option<&Tensor>,
+        cos: &Tensor,
+        sin: &Tensor,
+    ) -> Result<Tensor> {
+        anyhow::ensure!(
+            token_embed.dims() == hidden.dims(),
+            "mtp: embedding {:?} and hidden {:?} must have the same shape — \
+             the head fuses them position-for-position",
+            token_embed.dims(),
+            hidden.dims()
+        );
+        let e = self.pre_fc_norm_embedding.forward(token_embed)?;
+        let h = self.pre_fc_norm_hidden.forward(hidden)?;
+        // [embedding, hidden] — vLLM's order. See the module note.
+        let x = Tensor::cat(&[&e, &h], candle_core::D::Minus1)?;
+        let x = self.fc.forward(&x)?;
+        let x = self
+            .layer
+            .forward(&x, attn_mask, cos, sin)
+            .context("mtp draft layer")?;
+        Ok(self.norm.forward(&x)?)
+    }
+
+    /// Capture the draft layer's KV cache.
+    ///
+    /// Drafting advances the head over tokens that may be rejected, so
+    /// anything that drafts speculatively has to be able to put the
+    /// cache back. One full-attention layer is ~2 KiB/token/card, so
+    /// this is cheap next to the target's — which is why the observer
+    /// (#96 S4) can snapshot around every draft without distorting what
+    /// it measures.
+    pub fn snapshot_kv(&self) -> candle_core::Result<super::super::snapshot::LayerKvSnapshot> {
+        self.layer.snapshot_kv()
+    }
+
+    /// Put the draft layer's KV cache back, undoing a speculative walk.
+    pub fn restore_kv(
+        &mut self,
+        snap: &super::super::snapshot::LayerKvSnapshot,
+    ) -> candle_core::Result<()> {
+        self.layer.restore_kv(snap)
+    }
+
+    /// Drop the draft layer's KV cache. Independent of the target's.
+    pub fn clear_kv_cache(&mut self) {
+        self.layer.clear_kv_cache();
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use candle_core::{DType, Device};
+    use std::collections::HashMap;
+
+    /// Build a `ShardedVarBuilder` over a synthetic `mtp.*` head, the
+    /// way the sibling tests in this arch do: write a safetensors file
+    /// and open it.
+    fn tiny_head(
+        dir: &std::path::Path,
+        h: usize,
+        inter: usize,
+        heads: usize,
+        kv: usize,
+        head_dim: usize,
+    ) -> ShardedVarBuilder<'static> {
+        let dev = Device::Cpu;
+        let mut t: HashMap<String, Tensor> = HashMap::new();
+        let zeros = |shape: &[usize]| Tensor::zeros(shape, DType::F32, &dev).unwrap();
+        let randn = |shape: &[usize]| Tensor::randn(0f32, 0.1f32, shape, &dev).unwrap();
+
+        t.insert("mtp.fc.weight".into(), randn(&[h, 2 * h]));
+        t.insert("mtp.norm.weight".into(), zeros(&[h]));
+        t.insert("mtp.pre_fc_norm_embedding.weight".into(), zeros(&[h]));
+        t.insert("mtp.pre_fc_norm_hidden.weight".into(), zeros(&[h]));
+        let l = "mtp.layers.0";
+        t.insert(format!("{l}.input_layernorm.weight"), zeros(&[h]));
+        t.insert(format!("{l}.post_attention_layernorm.weight"), zeros(&[h]));
+        // Output-gated q_proj: 2 * heads * head_dim rows (attn_output_gate).
+        t.insert(
+            format!("{l}.self_attn.q_proj.weight"),
+            randn(&[2 * heads * head_dim, h]),
+        );
+        t.insert(
+            format!("{l}.self_attn.k_proj.weight"),
+            randn(&[kv * head_dim, h]),
+        );
+        t.insert(
+            format!("{l}.self_attn.v_proj.weight"),
+            randn(&[kv * head_dim, h]),
+        );
+        t.insert(
+            format!("{l}.self_attn.o_proj.weight"),
+            randn(&[h, heads * head_dim]),
+        );
+        t.insert(format!("{l}.self_attn.q_norm.weight"), zeros(&[head_dim]));
+        t.insert(format!("{l}.self_attn.k_norm.weight"), zeros(&[head_dim]));
+        t.insert(format!("{l}.mlp.gate_proj.weight"), randn(&[inter, h]));
+        t.insert(format!("{l}.mlp.up_proj.weight"), randn(&[inter, h]));
+        t.insert(format!("{l}.mlp.down_proj.weight"), randn(&[h, inter]));
+
+        let path = dir.join("mtp.safetensors");
+        candle_core::safetensors::save(&t, &path).expect("write tiny mtp checkpoint");
+        // SAFETY: mmaps a file this test just wrote and owns.
+        unsafe {
+            candle_nn::var_builder::ShardedSafeTensors::var_builder(
+                std::slice::from_ref(&path),
+                DType::F32,
+                &dev,
+            )
+        }
+        .expect("build ShardedVarBuilder")
+    }
+
+    fn tiny_cfg(h: usize, inter: usize, heads: usize, kv: usize, head_dim: usize) -> TextConfig {
+        let raw = format!(
+            r#"{{
+                "model_type": "qwen3_next",
+                "vocab_size": 32, "hidden_size": {h}, "intermediate_size": {inter},
+                "num_hidden_layers": 1, "num_attention_heads": {heads},
+                "num_key_value_heads": {kv}, "head_dim": {head_dim},
+                "max_position_embeddings": 64, "rms_norm_eps": 1e-6,
+                "attn_output_gate": true,
+                "layer_types": ["full_attention"],
+                "mtp_num_hidden_layers": 1,
+                "mtp_use_dedicated_embeddings": false
+            }}"#
+        );
+        super::super::Config::from_config_json(&raw)
+            .expect("parse tiny qwen3_5 config")
+            .text_config
+    }
+
+    /// The head loads from the `mtp.*` names the real checkpoints use
+    /// and produces finite output of the target's hidden width — which
+    /// is what the target's `lm_head` will consume.
+    #[test]
+    fn a_tiny_mtp_head_loads_and_drafts() {
+        let (h, inter, heads, kv, head_dim) = (8usize, 16usize, 2usize, 1usize, 4usize);
+        let dir = tempfile::tempdir().expect("tempdir");
+        let vb = tiny_head(dir.path(), h, inter, heads, kv, head_dim);
+        let cfg = tiny_cfg(h, inter, heads, kv, head_dim);
+        assert!(MtpHead::present_in(&cfg));
+
+        let dev = Device::Cpu;
+        let rotary = Arc::new(RotaryEmbedding::new(DType::F32, &cfg, &dev).expect("rotary"));
+        let mut head = MtpHead::load(&cfg, rotary.clone(), &vb).expect("load tiny mtp head");
+
+        let (b, l) = (1usize, 3usize);
+        let embed = Tensor::randn(0f32, 1.0, (b, l, h), &dev).unwrap();
+        let hidden = Tensor::randn(0f32, 1.0, (b, l, h), &dev).unwrap();
+        let positions: Vec<usize> = (0..l).collect();
+        let (cos, sin) = rotary.cos_sin_at(&positions).expect("cos/sin");
+        let out = head
+            .forward(&embed, &hidden, None, &cos, &sin)
+            .expect("draft step");
+        assert_eq!(out.dims(), &[b, l, h]);
+        let v: Vec<f32> = out.flatten_all().unwrap().to_vec1().unwrap();
+        assert!(
+            v.iter().all(|x| x.is_finite()),
+            "draft hidden must be finite"
+        );
+    }
+
+    /// `fc` consumes `cat[embedding, hidden]`, in that order.
+    ///
+    /// Both halves are `hidden_size` wide, so the reversed reading
+    /// loads, runs, and returns finite numbers that are simply a
+    /// different model. vLLM settles it —
+    /// `torch.cat([inputs_embeds, hidden_states], dim=-1)` — and this
+    /// pins it by value so a future edit cannot quietly swap them.
+    ///
+    /// The probe: `fc` reads only the embedding half (identity on the
+    /// first H columns, zero on the second). With a zero embedding and
+    /// a non-zero hidden state, the correct order yields zero and the
+    /// reversed order does not.
+    #[test]
+    fn the_fc_input_is_embedding_then_hidden() {
+        let dev = Device::Cpu;
+        let h = 4usize;
+        // fc = [I | 0] : takes the first H columns only.
+        let mut w = vec![0f32; h * 2 * h];
+        for i in 0..h {
+            w[i * 2 * h + i] = 1.0;
+        }
+        let fc = Linear::new(Tensor::from_vec(w, (h, 2 * h), &dev).unwrap(), None);
+        // Zero weight => (1 + w) = 1, a pure RMS normalisation, so the
+        // norms cannot mask the ordering.
+        let norm =
+            || Qwen3_5RmsNorm::from_weight(Tensor::zeros(h, DType::F32, &dev).unwrap(), 1e-6, None);
+
+        let e = Tensor::zeros((1, 1, h), DType::F32, &dev).unwrap();
+        let hid = Tensor::from_vec(vec![1f32, 2., 3., 4.], (1, 1, h), &dev).unwrap();
+
+        let en = norm().forward(&e).unwrap();
+        let hn = norm().forward(&hid).unwrap();
+
+        let correct: Vec<f32> = fc
+            .forward(&Tensor::cat(&[&en, &hn], candle_core::D::Minus1).unwrap())
+            .unwrap()
+            .flatten_all()
+            .unwrap()
+            .to_vec1()
+            .unwrap();
+        let reversed: Vec<f32> = fc
+            .forward(&Tensor::cat(&[&hn, &en], candle_core::D::Minus1).unwrap())
+            .unwrap()
+            .flatten_all()
+            .unwrap()
+            .to_vec1()
+            .unwrap();
+
+        assert!(
+            correct.iter().all(|x| x.abs() < 1e-6),
+            "with [embedding, hidden] the zero embedding must reach fc: {correct:?}"
+        );
+        assert!(
+            reversed.iter().any(|x| x.abs() > 1e-3),
+            "the reversed reading is what this test exists to catch: {reversed:?}"
+        );
+    }
+
+    /// `forward_multi` must agree with `forward` at the position they
+    /// share.
+    ///
+    /// The speculative verify pass (#96) reads the target's own token
+    /// at every drafted position, so the whole scheme rests on those
+    /// logits being the ones the model would have produced anyway.
+    ///
+    /// Agreement is to a tolerance, not bit-for-bit, and the reason is
+    /// worth recording: `forward` slices the last position *before* the
+    /// LM head and multiplies `(1, 1, H) x (H, vocab)`, while
+    /// `forward_multi` multiplies `(1, L, H) x (H, vocab)` and slices
+    /// after. Different GEMM shapes accumulate in a different order, so
+    /// the results differ in the last ULP or two — on this fixture,
+    /// 0.0016435327 against 0.0016435329. An exact assertion passed on
+    /// one CPU's BLAS blocking and failed on another's, which is the
+    /// only thing it was really testing.
+    ///
+    /// What matters for acceptance is that the argmax agrees, so that
+    /// is asserted too: a verify pass compares tokens, not floats.
+    #[test]
+    fn forward_multi_agrees_with_forward_at_the_last_position() {
+        use super::super::{Config, Qwen3_5ForCausalLM};
+        use candle_core::IndexOp;
+
+        let (h, inter, heads, kv, head_dim, vocab) =
+            (8usize, 16usize, 2usize, 1usize, 4usize, 32usize);
+        let dev = Device::Cpu;
+        let dir = tempfile::tempdir().expect("tempdir");
+        let mut t: HashMap<String, Tensor> = HashMap::new();
+        let zeros = |shape: &[usize]| Tensor::zeros(shape, DType::F32, &dev).unwrap();
+        let randn = |shape: &[usize]| Tensor::randn(0f32, 0.1f32, shape, &dev).unwrap();
+        t.insert("model.embed_tokens.weight".into(), randn(&[vocab, h]));
+        t.insert("lm_head.weight".into(), randn(&[vocab, h]));
+        t.insert("model.norm.weight".into(), zeros(&[h]));
+        let l = "model.layers.0";
+        t.insert(format!("{l}.input_layernorm.weight"), zeros(&[h]));
+        t.insert(format!("{l}.post_attention_layernorm.weight"), zeros(&[h]));
+        t.insert(
+            format!("{l}.self_attn.q_proj.weight"),
+            randn(&[2 * heads * head_dim, h]),
+        );
+        t.insert(
+            format!("{l}.self_attn.k_proj.weight"),
+            randn(&[kv * head_dim, h]),
+        );
+        t.insert(
+            format!("{l}.self_attn.v_proj.weight"),
+            randn(&[kv * head_dim, h]),
+        );
+        t.insert(
+            format!("{l}.self_attn.o_proj.weight"),
+            randn(&[h, heads * head_dim]),
+        );
+        t.insert(format!("{l}.self_attn.q_norm.weight"), zeros(&[head_dim]));
+        t.insert(format!("{l}.self_attn.k_norm.weight"), zeros(&[head_dim]));
+        t.insert(format!("{l}.mlp.gate_proj.weight"), randn(&[inter, h]));
+        t.insert(format!("{l}.mlp.up_proj.weight"), randn(&[inter, h]));
+        t.insert(format!("{l}.mlp.down_proj.weight"), randn(&[h, inter]));
+        let path = dir.path().join("model.safetensors");
+        candle_core::safetensors::save(&t, &path).expect("write tiny checkpoint");
+        // SAFETY: mmaps a file this test just wrote and owns.
+        let vb = unsafe {
+            candle_nn::var_builder::ShardedSafeTensors::var_builder(
+                std::slice::from_ref(&path),
+                DType::F32,
+                &dev,
+            )
+        }
+        .expect("build ShardedVarBuilder");
+
+        let raw = format!(
+            r#"{{
+                "model_type": "qwen3_next",
+                "vocab_size": {vocab}, "hidden_size": {h}, "intermediate_size": {inter},
+                "num_hidden_layers": 1, "num_attention_heads": {heads},
+                "num_key_value_heads": {kv}, "head_dim": {head_dim},
+                "max_position_embeddings": 64, "rms_norm_eps": 1e-6,
+                "attn_output_gate": true,
+                "layer_types": ["full_attention"]
+            }}"#
+        );
+        let cfg = Config::from_config_json(&raw).expect("parse tiny config");
+        let mut model = Qwen3_5ForCausalLM::new(cfg, vb).expect("load tiny checkpoint");
+
+        let input = Tensor::new(&[1u32, 5, 9, 2], &dev)
+            .unwrap()
+            .unsqueeze(0)
+            .unwrap();
+        let multi = model.forward_multi(&input, 0).expect("forward_multi");
+        assert_eq!(multi.dims(), &[1, 4, vocab], "one logits row per position");
+        let last_of_multi: Vec<f32> = multi.i((0, 3, ..)).unwrap().to_vec1().unwrap();
+        // A head that returned the same row L times would satisfy the
+        // comparison below without carrying per-position information,
+        // which is the whole point of the pass.
+        let first_of_multi: Vec<f32> = multi.i((0, 0, ..)).unwrap().to_vec1().unwrap();
+        assert_ne!(
+            first_of_multi, last_of_multi,
+            "positions must carry distinct logits, not a repeated row"
+        );
+
+        model.clear_kv_cache();
+        let single = model.forward(&input, 0).expect("forward");
+        assert_eq!(single.dims(), &[1, 1, vocab]);
+        let single: Vec<f32> = single.flatten_all().unwrap().to_vec1().unwrap();
+
+        let max_diff = last_of_multi
+            .iter()
+            .zip(single.iter())
+            .map(|(a, b)| (a - b).abs())
+            .fold(0f32, f32::max);
+        assert!(
+            max_diff < 1e-5,
+            "forward_multi's last row must match what forward returns \
+             (max abs diff {max_diff:e}): {last_of_multi:?} vs {single:?}"
+        );
+
+        let argmax = |v: &[f32]| {
+            v.iter()
+                .enumerate()
+                .max_by(|a, b| a.1.partial_cmp(b.1).expect("finite logits"))
+                .map(|(i, _)| i)
+                .expect("non-empty")
+        };
+        assert_eq!(
+            argmax(&last_of_multi),
+            argmax(&single),
+            "the verify pass compares tokens; the argmax must agree"
+        );
+    }
+}

crates/neuron/src/harness/arch/qwen4_exp/model.rs

@@ -45,6 +45,7 @@ use super::config::{Config, TextConfig};
 use super::decoder::DecoderLayer;
 use super::hyper::HyperConnection;
 use super::ple::{MmapNGramTable, NGramHasher, NGramTable, ShardedNGramTable};
+use crate::harness::arch::phase_probe::PhaseProbe;
 use crate::harness::arch::snapshot::{KvCacheSnapshot, PleSnapshot};
 
 /// The hashed n-gram lookup, hoisted to the model because it is a
@@ -255,6 +256,7 @@ impl Qwen4ExpForCausalLM {
 
         // Four identical streams — the residual this architecture
         // carries between layers, not one.
+        let mut probe = PhaseProbe::new(&self.device);
         let x = self.embed_tokens.forward(input_ids)?;
         let mut h = x.repeat((1, 1, self.hc_count))?;
 
@@ -264,6 +266,7 @@ impl Qwen4ExpForCausalLM {
         } else {
             None
         };
+        probe.mark("embed");
 
         // Computed once: it is a function of the ids, and one layer
         // consumes it.
@@ -274,12 +277,17 @@ impl Qwen4ExpForCausalLM {
             }
             None => None,
         };
+        // The n-gram table is mapped from the checkpoint rather than
+        // resident (#310), so this phase is a host page-fault path, not
+        // device work: it is where a cold or evicted mapping shows up.
+        probe.mark("ngram");
 
         for (i, layer) in self.layers.iter_mut().enumerate() {
             h = layer
                 .forward(&h, ngram.as_ref(), mask.as_ref(), &cos, &sin, offset)
                 .with_context(|| format!("layer {i}"))?;
         }
+        probe.mark("layers");
 
         // The mixer's hc_norm is the only normalisation before the head.
         let x = self.mixer.collapse(&h)?;
@@ -292,7 +300,10 @@ impl Qwen4ExpForCausalLM {
         // them all and [`Self::logits_all_positions`] exists to be
         // compared against it.
         let l = x.dim(1)?;
-        Ok((self.lm_head.forward(&x.i((.., l - 1.., ..))?)?, h))
+        let logits = self.lm_head.forward(&x.i((.., l - 1.., ..))?)?;
+        probe.mark("head");
+        probe.emit(seq_len);
+        Ok((logits, h))
     }
 
     /// Capture every piece of per-request state at one token boundary:

crates/neuron/src/harness/arch/qwen4_exp/ple.rs

@@ -420,6 +420,26 @@ impl MmapNGramTable {
             // defend against.
             let map = unsafe { memmap2::Mmap::map(&file) }
                 .with_context(|| format!("mmap {}", path.display()))?;
+            // Every access through this mapping is a 320-byte row picked
+            // by a hash: sixteen unrelated rows per token, out of
+            // 320,001,536. The kernel's default readahead assumes
+            // sequential access and faults a whole window per touch,
+            // which is pure amplification here -- and this table is
+            // 95 GiB against a host that is also holding the routed
+            // experts (#318) in RAM, so every wasted page evicts one
+            // that a later gather wanted.
+            //
+            // Advisory: a failure changes performance, never
+            // correctness, so it is logged rather than propagated. A
+            // kernel or filesystem that ignores MADV_RANDOM simply
+            // leaves the previous behaviour in place.
+            if let Err(e) = map.advise(memmap2::Advice::Random) {
+                tracing::debug!(
+                    file = %path.display(),
+                    error = %e,
+                    "ple: MADV_RANDOM refused; readahead stays on for this mapping"
+                );
+            }
             let (header_len, meta) =
                 safetensors::SafeTensors::read_metadata(&map).map_err(|e| {
                     anyhow::anyhow!("read safetensors metadata from {}: {e}", path.display())

crates/neuron/src/harness/arch/qwen4_exp/qsa.rs

@@ -56,6 +56,52 @@ use crate::harness::arch::qwen3_5::rope::{RotaryEmbedding, apply_partial_rotary}
 use super::config::TextConfig as Qwen4ExpTextConfig;
 
 /// The indexer's geometry, from `config.json`.
+/// The positions one query may attend, ascending.
+///
+/// The reference derives this from the attention mask
+/// (`torch.nonzero(visible_token_indices[b, 0, q])`) and forms blocks
+/// from whichever positions survive, in order. Under a plain causal
+/// mask that is exactly `0..n`, which is worth representing without
+/// materialising: the set reaches the context window in size and is
+/// rebuilt for every query, so a `Vec` per query would cost more than
+/// the selection it feeds.
+///
+/// [`Sparse`] is the general case — left padding, ragged batches, any
+/// mask with holes. Both are handled by the same arithmetic below,
+/// which is the point: block `b` covers *set positions*
+/// `b * block_size ..`, never *absolute positions*, and those coincide
+/// only when the set is contiguous from zero.
+///
+/// [`Sparse`]: VisibleSet::Sparse
+#[derive(Debug, Clone, Copy)]
+pub enum VisibleSet<'a> {
+    /// Every position `0..n` — a plain causal mask.
+    Contiguous(usize),
+    /// Exactly these positions, ascending and without duplicates.
+    Sparse(&'a [usize]),
+}
+
+impl VisibleSet<'_> {
+    pub fn len(&self) -> usize {
+        match self {
+            Self::Contiguous(n) => *n,
+            Self::Sparse(v) => v.len(),
+        }
+    }
+
+    pub fn is_empty(&self) -> bool {
+        self.len() == 0
+    }
+
+    /// The `i`th visible position.
+    fn at(&self, i: usize) -> usize {
+        match self {
+            Self::Contiguous(_) => i,
+            Self::Sparse(v) => v[i],
+        }
+    }
+}
+
 pub struct BlockSelector {
     n_heads: usize,
     head_dim: usize,
@@ -96,9 +142,14 @@ impl BlockSelector {
 
     /// The position each block starts at — where its pooled key is
     /// RoPE'd, **not** the position it is scored against.
-    pub fn block_starts(&self, n_visible: usize) -> Vec<usize> {
-        (0..self.n_blocks(n_visible))
-            .map(|b| b * self.block_size)
+    ///
+    /// Block `b` begins at the `b * block_size`-th **visible** position,
+    /// which is `b * block_size` only when the set is contiguous from
+    /// zero. Rotating a pooled key at the wrong absolute position is
+    /// silent: the scores stay finite and the selection stays plausible.
+    pub fn block_starts(&self, visible: &VisibleSet) -> Vec<usize> {
+        (0..self.n_blocks(visible.len()))
+            .map(|b| visible.at(b * self.block_size))
             .collect()
     }
 
@@ -149,7 +200,8 @@ impl BlockSelector {
     ///
     /// Ties are broken towards the earlier block, which matters only for
     /// reproducibility — the reference's `topk` does the same.
-    pub fn select(&self, scores: &[f32], n_visible: usize) -> Result<Vec<usize>> {
+    pub fn select(&self, scores: &[f32], visible: &VisibleSet) -> Result<Vec<usize>> {
+        let n_visible = visible.len();
         let blocks = self.n_blocks(n_visible);
         ensure!(
             scores.len() == blocks,
@@ -170,15 +222,28 @@ impl BlockSelector {
             order.sort_unstable();
         }
 
+        // Indices into the visible set, not absolute positions. They
+        // coincide only when the set is contiguous from zero, and
+        // conflating them is what this issue (#324) is about.
         let tail_start = blocks * self.block_size;
         let mut positions = Vec::with_capacity(keep * self.block_size + (n_visible - tail_start));
         for b in order {
             let start = b * self.block_size;
-            positions.extend(start..start + self.block_size);
+            positions.extend((start..start + self.block_size).map(|i| visible.at(i)));
         }
         // The tail never competes for the budget. It holds the query's
         // own position.
-        positions.extend(tail_start..n_visible);
+        positions.extend((tail_start..n_visible).map(|i| visible.at(i)));
+        // Ascending is the contract `dense_mask` relies on, and it
+        // holds without sorting: `order` is ascending, `visible.at` is
+        // monotonic in its index, and the tail's indices all follow the
+        // blocks'. Asserted rather than enforced, so a future change
+        // that breaks the invariant is caught instead of hidden by a
+        // sort that makes it look fine.
+        debug_assert!(
+            positions.windows(2).all(|w| w[0] < w[1]),
+            "selected positions must be ascending and distinct: {positions:?}"
+        );
         Ok(positions)
     }
 
@@ -356,7 +421,10 @@ impl Indexer {
         total: usize,
     ) -> candle_core::Result<Tensor> {
         let b = cache.dims()[0];
-        let starts = self.selector.block_starts(total);
+        // The cache holds every past position, so the poolable set is
+        // the contiguous 0..total. When a padded or ragged batch can
+        // reach here this becomes `Sparse` of that row's live positions.
+        let starts = self.selector.block_starts(&VisibleSet::Contiguous(total));
         let mut pooled = Vec::with_capacity(b);
         for row in 0..b {
             pooled.push(self.selector.pool(&cache.i(row)?)?);
@@ -475,10 +543,20 @@ impl Indexer {
             None => k,
         };
         let total = cache.dims()[1];
+        // This equality is also what makes the visible set contiguous:
+        // every query's set is built as `Contiguous(past_len + i + 1)`,
+        // which describes the batch only while one scalar `past_len`
+        // does. A ragged or left-padded batch presents a padded `total`
+        // that no single `past_len` matches, and trips this — which is
+        // the intent, because the alternative is a selection computed
+        // against the wrong past that still looks plausible (#324).
         ensure!(
             total == past_len + seq_len,
             "indexer cache holds {total} positions but the queries start at {past_len} \
-             and run for {seq_len} — the cache and the main KV have diverged"
+             and run for {seq_len} — the cache and the main KV have diverged, or this \
+             is a ragged/padded batch whose visible set is no longer contiguous. The \
+             latter needs the set derived from the attention mask (VisibleSet::Sparse) \
+             rather than from a count (#324)."
         );
         self.key_cache = Some(cache.clone());
         Ok((q, cache))
@@ -540,12 +618,16 @@ impl Indexer {
         for row in scores.iter().take(batch) {
             let mut per_query = Vec::with_capacity(seq_len);
             for (i, all_blocks) in row.iter().enumerate().take(seq_len) {
-                let visible = past_len + i + 1;
+                // Contiguous because `model.rs` builds the causal mask
+                // itself and no padding mask reaches this arch (#324).
+                // The moment one can, this is where the row's live
+                // positions are derived from it.
+                let visible = VisibleSet::Contiguous(past_len + i + 1);
                 // Only blocks wholly behind this query exist for it.
-                let visible_blocks = self.selector.n_blocks(visible);
+                let visible_blocks = self.selector.n_blocks(visible.len());
                 per_query.push(
                     self.selector
-                        .select(&all_blocks[..visible_blocks], visible)?,
+                        .select(&all_blocks[..visible_blocks], &visible)?,
                 );
             }
             out.push(per_query);
@@ -600,6 +682,63 @@ mod tests {
         assert_eq!(qsa.block_topk() * 4, 2048);
     }
 
+    /// Blocks are formed from consecutive **visible** positions, not
+    /// consecutive absolute ones (#324).
+    ///
+    /// The reference takes `nonzero` of the mask and chunks whatever
+    /// survives, in order. With holes in the visible set the two
+    /// readings diverge completely: under the old arithmetic the query
+    /// would be handed positions 4..8, which here are not merely the
+    /// wrong block — 4, 6, 8, 9 and 10 are positions the mask forbids
+    /// it from seeing at all.
+    #[test]
+    fn blocks_are_cut_from_the_visible_set_not_the_position_axis() {
+        // block_size 4, budget 4 => one block survives the top-k.
+        let qsa = BlockSelector::new(2, 1, 4, 4).unwrap();
+        let live = [3usize, 5, 7, 11, 13, 17, 19, 23, 29];
+        let visible = VisibleSet::Sparse(&live);
+
+        assert_eq!(qsa.n_blocks(visible.len()), 2);
+        // Block 0 begins at visible[0] = 3, block 1 at visible[4] = 13.
+        assert_eq!(qsa.block_starts(&visible), vec![3, 13]);
+
+        // Score block 1 above block 0, so only block 1 is kept.
+        let got = qsa.select(&[0.1f32, 0.9], &visible).unwrap();
+        assert_eq!(
+            got,
+            vec![13, 17, 19, 23, 29],
+            "block 1 is visible[4..8], and the tail is visible[8..]"
+        );
+        // The failure this pins: index arithmetic would give 4..8 plus 8.
+        assert!(
+            !got.contains(&4) && !got.contains(&8),
+            "positions the mask forbids must never be selected: {got:?}"
+        );
+    }
+
+    /// The two representations must agree wherever both are valid.
+    ///
+    /// `Contiguous(n)` is an optimisation for the causal case, not a
+    /// different rule — if it can disagree with the explicit list of the
+    /// same positions, one of the two paths is wrong and only the
+    /// explicit one is checked against the reference.
+    #[test]
+    fn contiguous_and_sparse_agree_on_the_same_positions() {
+        let qsa = BlockSelector::new(2, 1, 4, 8).unwrap();
+        for n in [1usize, 4, 7, 8, 9, 15, 23] {
+            let explicit: Vec<usize> = (0..n).collect();
+            let scores = vec![0.5f32; qsa.n_blocks(n)];
+            let a = qsa.select(&scores, &VisibleSet::Contiguous(n)).unwrap();
+            let b = qsa.select(&scores, &VisibleSet::Sparse(&explicit)).unwrap();
+            assert_eq!(a, b, "representations disagree at n={n}");
+            assert_eq!(
+                qsa.block_starts(&VisibleSet::Contiguous(n)),
+                qsa.block_starts(&VisibleSet::Sparse(&explicit)),
+                "block starts disagree at n={n}"
+            );
+        }
+    }
+
     /// The cheapest gate in the architecture: below the budget, the
     /// selection returns every visible position, so a short prompt must
     /// be bit-comparable to dense causal attention.
@@ -609,7 +748,9 @@ mod tests {
         for n_visible in [1usize, 3, 4, 7, 2047, 2048] {
             assert!(qsa.is_dense(n_visible), "{n_visible} should be dense");
             let scores = vec![0.0f32; qsa.n_blocks(n_visible)];
-            let got = qsa.select(&scores, n_visible).unwrap();
+            let got = qsa
+                .select(&scores, &VisibleSet::Contiguous(n_visible))
+                .unwrap();
             assert_eq!(
                 got,
                 (0..n_visible).collect::<Vec<_>>(),
@@ -631,7 +772,9 @@ mod tests {
         assert_eq!(qsa.n_blocks(n_visible), 4);
         // Block 2 is worthless; with topk 3 it is the one dropped.
         let scores = [0.9f32, 0.8, 0.1, 0.7];
-        let got = qsa.select(&scores, n_visible).unwrap();
+        let got = qsa
+            .select(&scores, &VisibleSet::Contiguous(n_visible))
+            .unwrap();
         let want: Vec<usize> = (0..8).chain(12..19).collect();
         assert_eq!(got, want);
         assert!(got.contains(&(n_visible - 1)), "the query's own position");
@@ -721,7 +864,7 @@ mod tests {
         );
         // And the selection that rests on it picks the right block.
         assert_eq!(
-            qsa.select(&scores, 8).unwrap(),
+            qsa.select(&scores, &VisibleSet::Contiguous(8)).unwrap(),
             vec![4, 5, 6, 7],
             "the higher-scoring block should be the one kept"
         );
@@ -755,24 +898,26 @@ mod tests {
         let qsa = small();
         let keys = Tensor::zeros((11, 8), DType::BF16, &dev).unwrap();
         assert_eq!(qsa.pool(&keys).unwrap().dims(), &[2, 8]);
-        assert_eq!(qsa.block_starts(11), vec![0, 4]);
+        assert_eq!(qsa.block_starts(&VisibleSet::Contiguous(11)), vec![0, 4]);
 
         let short = Tensor::zeros((3, 8), DType::F32, &dev).unwrap();
         assert_eq!(qsa.pool(&short).unwrap().dims(), &[0, 8]);
-        assert!(qsa.block_starts(3).is_empty());
+        assert!(qsa.block_starts(&VisibleSet::Contiguous(3)).is_empty());
     }
 
     #[test]
     fn selection_is_ascending_and_rejects_a_wrong_score_count() {
         let qsa = small();
-        let got = qsa.select(&[0.1f32, 0.9, 0.5, 0.7, 0.2], 23).unwrap();
+        let got = qsa
+            .select(&[0.1f32, 0.9, 0.5, 0.7, 0.2], &VisibleSet::Contiguous(23))
+            .unwrap();
         assert!(got.windows(2).all(|w| w[0] < w[1]), "got {got:?}");
         // blocks 1, 3, 2 win; block 0 and 4 lose; tail 20..23 stays.
         assert_eq!(
             got,
             vec![4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 20, 21, 22]
         );
-        assert!(qsa.select(&[0.1f32], 23).is_err());
+        assert!(qsa.select(&[0.1f32], &VisibleSet::Contiguous(23)).is_err());
     }
 
     // ---- the indexer as a layer ----
@@ -832,6 +977,61 @@ mod tests {
         .unwrap()
     }
 
+    /// A cache that does not match `past_len + seq_len` is a ragged or
+    /// left-padded batch, and must fail loudly (#324).
+    ///
+    /// Every visible set is built as `Contiguous(past_len + i + 1)`, so
+    /// the moment one scalar `past_len` stops describing every row, the
+    /// selection is computed against the wrong past — and it stays
+    /// plausible, which is what makes it dangerous. The guard is only
+    /// worth having if it can fire, so this drives it directly rather
+    /// than trusting that it would.
+    #[test]
+    fn a_cache_inconsistent_with_past_len_is_refused() {
+        let (hidden_size, n_heads, head_dim) = (8, 2, 8);
+        let mut ix = indexer(hidden_size, n_heads, head_dim, 4, 4);
+        let rope = rope_for(head_dim);
+
+        // A history of 20 positions, while the step claims past_len 5:
+        // the shape a padded batch presents, where `total` is the
+        // padded maximum and no single `past_len` matches every row.
+        let stale = Tensor::zeros((1, 20, head_dim), DType::F32, &Device::Cpu).unwrap();
+        ix.restore_keys(Some(&stale));
+
+        let err = ix
+            .visible_positions(&hidden(1, 1, hidden_size, 0.7), &rope, 5)
+            .expect_err("a cache that disagrees with past_len must not be served");
+        let msg = format!("{err:#}");
+        assert!(
+            msg.contains("no longer contiguous") && msg.contains("324"),
+            "the error must name the contiguity assumption it protects, got: {msg}"
+        );
+    }
+
+    /// The same call with a consistent cache must be served, or the
+    /// guard above is just refusing everything.
+    #[test]
+    fn a_cache_consistent_with_past_len_is_served() {
+        let (hidden_size, n_heads, head_dim) = (8, 2, 8);
+        let mut ix = indexer(hidden_size, n_heads, head_dim, 4, 4);
+        let rope = rope_for(head_dim);
+
+        let history = Tensor::zeros((1, 5, head_dim), DType::F32, &Device::Cpu).unwrap();
+        ix.restore_keys(Some(&history));
+
+        // past_len 5 + this step's 1 == the 6 the cache will hold.
+        let got = ix
+            .visible_positions(&hidden(1, 1, hidden_size, 0.7), &rope, 5)
+            .expect("a contiguous history must still be served");
+        assert_eq!(got.len(), 1, "one batch row");
+        assert_eq!(got[0].len(), 1, "one query");
+        assert!(
+            got[0][0].iter().all(|p| *p < 6),
+            "positions must lie inside the 6-position history: {:?}",
+            got[0][0]
+        );
+    }
+
     /// `index_qk_proj` is fused 4:1, not halved. With 4 query heads and
     /// 1 key head of 8 dims the split is 32 and 8 of 40 — a halving
     /// would take 20 and 20 and put half a query head in the key.

crates/neuron/src/harness/arch/snapshot.rs

@@ -391,6 +391,8 @@ mod tests {
     fn tiny_config() -> TextConfig {
         TextConfig {
             vocab_size: 32,
+            mtp_num_hidden_layers: 0,
+            mtp_use_dedicated_embeddings: false,
             hidden_size: 16,
             intermediate_size: 32,
             num_hidden_layers: 2,

crates/neuron/src/harness/candle.rs

@@ -25,6 +25,11 @@ use cortex_core::openai::{
     ChatCompletionChunk, ChatCompletionRequest, ChatCompletionResponse, ChatMessage, MessageContent,
 };
 
+// `PhaseTiming` is constructed only on the two cuda-gated streaming
+// paths (device-worker and TP), so importing it unconditionally warns
+// on a CPU-only build and clippy runs with `-D warnings`.
+#[cfg(feature = "cuda")]
+use crate::wire::PhaseTiming;
 use crate::wire::{
     FinishReason, FinishTiming, InferenceEvent, ReasoningTokenPair, ToolCallTokenPair,
     detect_reasoning_token_pair, detect_tool_call_token_pair, openai_chat as wire_chat,
@@ -138,6 +143,9 @@ pub enum LoadedHandle {
     Tp(Arc<TpLoadedModel>),
     /// Z-Image text-to-image pipeline (#198).
     Image(Arc<LoadedImageModel>),
+    /// Laya decision model (#336): typed questions in, calibrated
+    /// option scores out, no text generation.
+    Decision(Arc<super::decision_model::LoadedDecisionModel>),
 }
 
 impl LoadedHandle {
@@ -147,6 +155,7 @@ impl LoadedHandle {
             #[cfg(feature = "cuda")]
             LoadedHandle::Tp(m) => &m.model_id,
             LoadedHandle::Image(m) => &m.model_id,
+            LoadedHandle::Decision(m) => &m.model_id,
         }
     }
 
@@ -157,6 +166,7 @@ impl LoadedHandle {
             #[cfg(feature = "cuda")]
             LoadedHandle::Tp(m) => &m.spec,
             LoadedHandle::Image(m) => &m.spec,
+            LoadedHandle::Decision(m) => &m.spec,
         }
     }
 
@@ -166,6 +176,7 @@ impl LoadedHandle {
             #[cfg(feature = "cuda")]
             LoadedHandle::Tp(m) => m.devices.clone(),
             LoadedHandle::Image(m) => m.devices.clone(),
+            LoadedHandle::Decision(m) => m.devices.clone(),
         }
     }
 
@@ -179,6 +190,7 @@ impl LoadedHandle {
             #[cfg(feature = "cuda")]
             LoadedHandle::Tp(m) => m.poisoned.load(Ordering::Acquire),
             LoadedHandle::Image(m) => m.poisoned.load(Ordering::Acquire),
+            LoadedHandle::Decision(m) => m.poisoned.load(Ordering::Acquire),
         }
     }
 
@@ -190,6 +202,7 @@ impl LoadedHandle {
             #[cfg(feature = "cuda")]
             LoadedHandle::Tp(m) => (m.admission.in_flight(), m.admission.queue_depth()),
             LoadedHandle::Image(m) => (m.admission.in_flight(), m.admission.queue_depth()),
+            LoadedHandle::Decision(m) => (m.admission.in_flight(), m.admission.queue_depth()),
         }
     }
 
@@ -202,6 +215,9 @@ impl LoadedHandle {
             #[cfg(feature = "cuda")]
             LoadedHandle::Tp(m) => (m.admission.max_in_flight(), m.admission.max_queue_depth()),
             LoadedHandle::Image(m) => (m.admission.max_in_flight(), m.admission.max_queue_depth()),
+            LoadedHandle::Decision(m) => {
+                (m.admission.max_in_flight(), m.admission.max_queue_depth())
+            }
         }
     }
 
@@ -225,6 +241,10 @@ impl LoadedHandle {
                 m.admission.anon_in_flight(),
                 m.admission.anon_max_in_flight(),
             ),
+            LoadedHandle::Decision(m) => (
+                m.admission.anon_in_flight(),
+                m.admission.anon_max_in_flight(),
+            ),
         }
     }
 
@@ -235,6 +255,7 @@ impl LoadedHandle {
             #[cfg(feature = "cuda")]
             LoadedHandle::Tp(m) => m.admission.close(),
             LoadedHandle::Image(m) => m.admission.close(),
+            LoadedHandle::Decision(m) => m.admission.close(),
         }
     }
 
@@ -246,6 +267,7 @@ impl LoadedHandle {
             #[cfg(feature = "cuda")]
             LoadedHandle::Tp(m) => m.admission.rejections(),
             LoadedHandle::Image(m) => m.admission.rejections(),
+            LoadedHandle::Decision(m) => m.admission.rejections(),
         }
     }
 
@@ -257,6 +279,7 @@ impl LoadedHandle {
             #[cfg(feature = "cuda")]
             LoadedHandle::Tp(m) => &m.admission,
             LoadedHandle::Image(m) => &m.admission,
+            LoadedHandle::Decision(m) => &m.admission,
         };
         (admission.kv_budget_mb(), admission.kv_available_mb())
     }
@@ -270,7 +293,7 @@ impl LoadedHandle {
             LoadedHandle::Tp(m) => &m.prefill_rate,
             // Image models have no token throughput; steps/sec lives in
             // the per-request timing instead.
-            LoadedHandle::Image(_) => return (0.0, 0.0),
+            LoadedHandle::Image(_) | LoadedHandle::Decision(_) => return (0.0, 0.0),
         };
         (ema.get().unwrap_or(0.0), ema.decode().unwrap_or(0.0))
     }
@@ -282,12 +305,15 @@ impl LoadedHandle {
         if matches!(self, LoadedHandle::Image(_)) {
             return vec!["image".to_string()];
         }
+        if matches!(self, LoadedHandle::Decision(_)) {
+            return vec!["decision".to_string()];
+        }
         let mut caps = vec!["text".to_string()];
         let has_vision = match self {
             LoadedHandle::Single(m) => m.has_vision,
             #[cfg(feature = "cuda")]
             LoadedHandle::Tp(m) => m.has_vision,
-            LoadedHandle::Image(_) => unreachable!("handled above"),
+            LoadedHandle::Image(_) | LoadedHandle::Decision(_) => unreachable!("handled above"),
         };
         if has_vision {
             caps.push("vision".to_string());
@@ -333,7 +359,7 @@ impl LoadedHandle {
             LoadedHandle::Tp(m) => (m.last_free_mb.load(Ordering::Acquire), m.context_profile),
             // Image generation has its own VRAM path and no prefill
             // floor, so this check would say nothing useful about it.
-            LoadedHandle::Image(_) => return None,
+            LoadedHandle::Image(_) | LoadedHandle::Decision(_) => return None,
         };
         // `0` is the CPU-load and missing-worker sentinel — the same one
         // `check_vram` skips on. Nothing to assert either way.
@@ -414,7 +440,7 @@ impl LoadedHandle {
             LoadedHandle::Single(m) => m.tool_call_tokens.is_some(),
             #[cfg(feature = "cuda")]
             LoadedHandle::Tp(m) => m.tool_call_tokens.is_some(),
-            LoadedHandle::Image(_) => false,
+            LoadedHandle::Image(_) | LoadedHandle::Decision(_) => false,
         }
     }
 
@@ -425,7 +451,7 @@ impl LoadedHandle {
             LoadedHandle::Single(m) => m.reasoning_tokens.is_some(),
             #[cfg(feature = "cuda")]
             LoadedHandle::Tp(m) => m.reasoning_tokens.is_some(),
-            LoadedHandle::Image(_) => false,
+            LoadedHandle::Image(_) | LoadedHandle::Decision(_) => false,
         }
     }
 
@@ -436,7 +462,7 @@ impl LoadedHandle {
             LoadedHandle::Single(m) => m.preserve_thinking,
             #[cfg(feature = "cuda")]
             LoadedHandle::Tp(m) => m.preserve_thinking,
-            LoadedHandle::Image(_) => None,
+            LoadedHandle::Image(_) | LoadedHandle::Decision(_) => None,
         }
     }
 
@@ -446,7 +472,7 @@ impl LoadedHandle {
             LoadedHandle::Single(m) => m.reasoning_efforts.clone(),
             #[cfg(feature = "cuda")]
             LoadedHandle::Tp(m) => m.reasoning_efforts.clone(),
-            LoadedHandle::Image(_) => Default::default(),
+            LoadedHandle::Image(_) | LoadedHandle::Decision(_) => Default::default(),
         }
     }
 
@@ -472,7 +498,7 @@ impl LoadedHandle {
             #[cfg(feature = "cuda")]
             LoadedHandle::Tp(m) => m.query_vram_tightest_free_mb().await,
             // No context-limit derivation for image models; nothing to cache.
-            LoadedHandle::Image(_) => return,
+            LoadedHandle::Image(_) | LoadedHandle::Decision(_) => return,
         };
         // Don't clobber a good cached value with a transient `0`
         // (worker gone/poisoned sentinel).
@@ -481,7 +507,7 @@ impl LoadedHandle {
                 LoadedHandle::Single(m) => m.last_free_mb.store(free, Ordering::Release),
                 #[cfg(feature = "cuda")]
                 LoadedHandle::Tp(m) => m.last_free_mb.store(free, Ordering::Release),
-                LoadedHandle::Image(_) => {}
+                LoadedHandle::Image(_) | LoadedHandle::Decision(_) => {}
             }
         }
     }
@@ -510,7 +536,7 @@ impl LoadedHandle {
                 m.prefill_rate.get(),
             ),
             // Token limits are meaningless for image generation.
-            LoadedHandle::Image(_) => return None,
+            LoadedHandle::Image(_) | LoadedHandle::Decision(_) => return None,
         };
         let rate = rate.unwrap_or(cfg.bootstrap_prefill_tok_per_sec);
         // The KV budget admission will actually grant. Advertising past
@@ -520,7 +546,9 @@ impl LoadedHandle {
             LoadedHandle::Single(m) => m.admission.kv_budget_mb(),
             #[cfg(feature = "cuda")]
             LoadedHandle::Tp(m) => m.admission.kv_budget_mb(),
-            LoadedHandle::Image(_) => unreachable!("derived_limit returned above"),
+            LoadedHandle::Image(_) | LoadedHandle::Decision(_) => {
+                unreachable!("derived_limit returned above")
+            }
         };
         let limit = super::context_limit::derive_limit(
             &profile,
@@ -539,7 +567,9 @@ impl LoadedHandle {
                 LoadedHandle::Single(m) => &m.derived_input_cap,
                 #[cfg(feature = "cuda")]
                 LoadedHandle::Tp(m) => &m.derived_input_cap,
-                LoadedHandle::Image(_) => unreachable!("derived_limit returned above"),
+                LoadedHandle::Image(_) | LoadedHandle::Decision(_) => {
+                    unreachable!("derived_limit returned above")
+                }
             };
             cap.store(input, Ordering::Release);
         }
@@ -1162,6 +1192,18 @@ impl ModelArch {
         }
     }
 
+    /// Logits at every position of a multi-token forward: `(B, L,
+    /// vocab)`. The speculative verify pass (#96) reads the target's
+    /// own token at each drafted position; `squeeze_to_vocab` is
+    /// deliberately not applied, as it would collapse the L axis this
+    /// exists to keep.
+    pub fn forward_multi(&mut self, input: &Tensor, offset: usize) -> Result<Tensor> {
+        match self {
+            ModelArch::Qwen3_5Dense(m) => Ok(m.forward_multi(input, offset)?),
+            _ => anyhow::bail!("forward_multi: architecture has no multi-position logits support"),
+        }
+    }
+
     /// Padding mask for a batched decode step — see
     /// `Qwen3_5Model::batch_decode_mask`.
     pub fn batch_decode_mask(
@@ -3265,7 +3307,7 @@ impl CandleHarness {
             LoadedHandle::Tp(m) => {
                 return self.inference_tp_stream(m, request, principal, wire).await;
             }
-            LoadedHandle::Image(_) => {
+            LoadedHandle::Image(_) | LoadedHandle::Decision(_) => {
                 return Err(InferenceError::WrongModality {
                     model_id: request.model.clone(),
                 });
@@ -3505,11 +3547,13 @@ impl CandleHarness {
                 let reasoning_tokens_inner = loaded.reasoning_tokens.clone();
                 let tool_call_tokens_inner = loaded.tool_call_tokens.clone();
                 let tool_schemas_inner = tool_schemas.clone();
+                let model_id_inner = model_id.clone();
                 tokio::spawn(
                     async move {
                         let _admit = admit;
                         let _inference_guard = loaded_for_task.inference_lock.lock().await;
                         match stream_inference_via_worker(
+                            model_id_inner,
                             worker,
                             handle,
                             tokenizer,
@@ -3982,6 +4026,11 @@ impl Harness for CandleHarness {
         if matches!(plan.format, super::preflight::SourceFormat::Diffusers) {
             return self.load_image_model(spec, &source_id).await;
         }
+        // Laya decision models (#336) take the decision path: one
+        // device, unquantized (preflight rejected tp>1 and quant).
+        if matches!(plan.format, super::preflight::SourceFormat::DecisionModel) {
+            return self.load_decision_model(spec, &source_id).await;
+        }
 
         let tp_size = spec.tensor_parallel.unwrap_or(1);
         if tp_size > 1 {
@@ -4348,6 +4397,11 @@ impl Harness for CandleHarness {
                     );
                 }
             }
+            LoadedHandle::Decision(dm) => {
+                // Each checkpoint's weights drop on the thread that owns
+                // them (the device worker, for a CUDA load).
+                dm.release().await;
+            }
             #[cfg(feature = "cuda")]
             LoadedHandle::Tp(tp) => {
                 // Try to recover the inner TpLoadedModel so we can move
@@ -4522,6 +4576,132 @@ impl CandleHarness {
         Ok(())
     }
 
+    /// Load a Laya decision model (#336): resolve the checkpoints' files,
+    /// build them on the device worker (or the CPU), and register the
+    /// model with `capabilities: ["decision"]`.
+    ///
+    /// The repo root is required. A released family also carries
+    /// `multilingual/` and `typed-decisions/` beside it (#339); each one
+    /// the repo has is loaded too, so the `/v1/systemone` handler can
+    /// route non-English states to the multilingual checkpoint and honour
+    /// a request that pins one. A repo without them (a single fine-tune)
+    /// serves its root alone.
+    async fn load_decision_model(
+        &self,
+        spec: &ModelSpec,
+        source_id: &cortex_core::source::ModelSourceId,
+    ) -> Result<()> {
+        let devices = spec.devices.clone().unwrap_or_else(|| vec![0]);
+        let device = Self::pick_device(&devices)?;
+        #[cfg(feature = "cuda")]
+        let worker = if matches!(device, Device::Cuda(_)) {
+            Some(self.ensure_device_worker(devices[0]).await?)
+        } else {
+            None
+        };
+        // CPU build: always the CPU backend. Correct, but at hundreds of
+        // milliseconds per question it is a fallback, not a serving tier.
+        #[cfg(not(feature = "cuda"))]
+        let worker = {
+            let _ = &device;
+            None
+        };
+
+        let files = self.resolve_decision_files(source_id, None).await?;
+        let root = super::decision_model::DecisionCheckpoint::load(
+            super::decision_model::ROOT_CHECKPOINT.to_string(),
+            files,
+            &spec.model_id,
+            worker.clone(),
+        )
+        .await?;
+        let dtype = root.dtype;
+        let mut checkpoints = vec![root];
+        for sub in super::decision::route::FAMILY
+            .iter()
+            .filter(|c| **c != super::decision_model::ROOT_CHECKPOINT)
+        {
+            let files = match self.resolve_decision_files(source_id, Some(sub)).await {
+                Ok(files) => files,
+                Err(e) => {
+                    tracing::info!(
+                        model = %spec.model_id,
+                        checkpoint = sub,
+                        error = %format!("{e:#}"),
+                        "decision checkpoint not available; serving without it"
+                    );
+                    continue;
+                }
+            };
+            checkpoints.push(
+                super::decision_model::DecisionCheckpoint::load(
+                    sub.to_string(),
+                    files,
+                    &spec.model_id,
+                    worker.clone(),
+                )
+                .await?,
+            );
+        }
+        let names: Vec<&str> = checkpoints.iter().map(|c| c.name.as_str()).collect();
+        tracing::info!(model = %spec.model_id, checkpoints = ?names, "decision checkpoints loaded");
+
+        let loaded = super::decision_model::LoadedDecisionModel {
+            model_id: spec.model_id.clone(),
+            spec: spec.clone(),
+            devices,
+            checkpoints,
+            poisoned: Arc::new(AtomicBool::new(false)),
+            admission: crate::harness::admission::AdmissionController::new(&self.admission_cfg),
+        };
+        let mut models = self.models.write().await;
+        models.insert(
+            spec.model_id.clone(),
+            LoadedHandle::Decision(Arc::new(loaded)),
+        );
+        tracing::info!(model = %spec.model_id, dtype = ?dtype, "decision model loaded");
+        Ok(())
+    }
+
+    /// Fetch one Laya checkpoint's files through hf-hub: the repo root
+    /// for `subfolder = None`, else that subfolder. Only the files the
+    /// checkpoint needs are fetched, not the whole repo.
+    async fn resolve_decision_files(
+        &self,
+        source_id: &cortex_core::source::ModelSourceId,
+        subfolder: Option<&str>,
+    ) -> Result<super::decision_model::CheckpointFiles> {
+        let api = self.hf_api_for(&source_id.scheme)?;
+        let repo = api.model(source_id.repo_path());
+        let display_id = source_id.to_string();
+        let prefix = subfolder.map(|s| format!("{s}/")).unwrap_or_default();
+        let mut fetched = Vec::new();
+        for file in [
+            "rl_agent_config.json",
+            "encoder/config.json",
+            "model.safetensors",
+            "tokenizer/tokenizer.json",
+            "tokenizer/tokenizer_config.json",
+        ] {
+            let path = format!("{prefix}{file}");
+            fetched.push(
+                repo.get(&path)
+                    .await
+                    .with_context(|| format!("fetch {path} from {display_id}"))?,
+            );
+        }
+        // hf-hub materialises a snapshot as a directory tree mirroring
+        // the repo, so the checkpoint directory is where its config sits.
+        let dir = fetched[0]
+            .parent()
+            .context("checkpoint config has no parent directory")?
+            .to_path_buf();
+        Ok(super::decision_model::CheckpointFiles {
+            dir,
+            tokenizer: fetched[3].clone(),
+        })
+    }
+
     /// Resolve a diffusers-layout Z-Image repo into local component
     /// paths, downloading through hf-hub as needed. Verifies the
     /// pipeline class before fetching multi-GB weight shards.
@@ -4635,6 +4815,189 @@ impl CandleHarness {
         }
     }
 
+    /// The loaded decision model `model_id` (#336), for the
+    /// `/v1/systemone` handler: it takes the admission permit, scores
+    /// through [`super::decision_model::DecisionCheckpoint::decide`],
+    /// and shapes the answers. A model of another kind is a
+    /// `WrongModality`, as for images.
+    pub async fn decision_model(
+        &self,
+        model_id: &str,
+    ) -> Result<Arc<super::decision_model::LoadedDecisionModel>, InferenceError> {
+        let handle = {
+            let models = self.models.read().await;
+            models.get(model_id).cloned()
+        };
+        match handle {
+            Some(LoadedHandle::Decision(m)) => Ok(m),
+            Some(_) => Err(InferenceError::WrongModality {
+                model_id: model_id.to_string(),
+            }),
+            None => Err(InferenceError::ModelNotLoaded(model_id.to_string())),
+        }
+    }
+
+    /// Answer one `/v1/systemone` request (#338): pick the served decision
+    /// model and the checkpoint of it that answers, assemble the question
+    /// sequences, score them in one forward pass, and shape the answers.
+    ///
+    /// The served model is `served` when the gateway routed the request
+    /// (the `x-helexa-decision-model` header), else the model the body's
+    /// `model` names if one is loaded, else the one decision model this
+    /// node serves. The body's `model` is then read a second time, as the
+    /// reference server reads it: as an optional pin to a checkpoint of
+    /// the family (`multilingual`, `typed-decisions`, …).
+    ///
+    /// Questions are validated and tokenised before admission, so a
+    /// malformed request never takes a slot.
+    pub async fn systemone(
+        &self,
+        served: Option<&str>,
+        request: cortex_core::decisions::SystemOneRequest,
+        principal: Option<&str>,
+    ) -> Result<cortex_core::decisions::SystemOneResponse, SystemOneError> {
+        use super::decision::{self, route};
+
+        let model = match served {
+            Some(id) => self.decision_model(id).await?,
+            None => {
+                self.default_decision_model(request.model.as_deref())
+                    .await?
+            }
+        };
+        if model.poisoned.load(Ordering::Acquire) {
+            return Err(InferenceError::Other(anyhow::anyhow!(
+                "decision model '{}' is poisoned; unload and reload it",
+                model.model_id
+            ))
+            .into());
+        }
+
+        let pin = route::resolve_pin(request.model.as_deref());
+        let mut routing = route::route(&model.model_id, &request.state, &request.questions, pin);
+        let checkpoint = match model.checkpoint(Some(&routing.model)) {
+            Some(c) => c,
+            None if pin.is_some() => {
+                return Err(decision::DecisionRequestError::new(
+                    404,
+                    format!(
+                        "checkpoint '{}' is not served by '{}' on this node",
+                        routing.model, model.model_id
+                    ),
+                )
+                .into());
+            }
+            // Routed (not pinned) to a checkpoint this repo does not
+            // carry: answer with the default one, and say so.
+            None => {
+                let fallback = model.checkpoint(None).ok_or_else(|| {
+                    InferenceError::Other(anyhow::anyhow!(
+                        "decision model '{}' has no checkpoints",
+                        model.model_id
+                    ))
+                })?;
+                routing.reason = format!(
+                    "{}; the {} checkpoint is not loaded, so {} answered",
+                    routing.reason, routing.model, fallback.name
+                );
+                routing.model = fallback.name.clone();
+                routing.repo = route::repo_of(&model.model_id, &fallback.name);
+                fallback
+            }
+        };
+        let lang = route::detected_language(&routing).map(str::to_string);
+
+        let prepared = decision::prepare(
+            &request.state,
+            &request.questions,
+            &checkpoint.calibration,
+            &checkpoint.tokenizer,
+            decision::Budget {
+                max_len: request.max_len,
+                head_max_len: request.head_max_len,
+            },
+        )?;
+
+        let _permit = model.admission.enter(principal).await?;
+        let outputs: Vec<decision::RowOutput> = if prepared.rows.is_empty() {
+            Vec::new()
+        } else {
+            let batch = super::arch::laya::DecisionBatch {
+                seqs: prepared
+                    .rows
+                    .iter()
+                    .map(|r| super::arch::laya::DecisionSeq {
+                        input_ids: r.input_ids.clone(),
+                        markers: r.markers.clone(),
+                        qtype: r.qtype as u8,
+                    })
+                    .collect(),
+            };
+            checkpoint
+                .decide(batch)
+                .await
+                .map_err(|e| InferenceError::Other(anyhow::anyhow!("decision forward: {e:#}")))?
+                .into_iter()
+                .map(|row| decision::RowOutput {
+                    act_probability: row.act.first().copied().unwrap_or(0.0),
+                    logits: row.logits,
+                })
+                .collect()
+        };
+        let (answers, usage) = decision::shape(
+            &prepared,
+            &outputs,
+            &checkpoint.calibration,
+            lang.as_deref(),
+        );
+        Ok(decision::shape::response(answers, usage, Some(routing)))
+    }
+
+    /// Register an already-built decision model, bypassing the hub fetch
+    /// `load_model` does — for tests that serve local checkpoints.
+    #[cfg(test)]
+    pub(crate) async fn register_decision_model(
+        &self,
+        model: super::decision_model::LoadedDecisionModel,
+    ) {
+        self.models.write().await.insert(
+            model.model_id.clone(),
+            LoadedHandle::Decision(Arc::new(model)),
+        );
+    }
+
+    /// The decision model to serve a request no gateway routed: the one
+    /// `requested` names if it is loaded, else the only (first) loaded
+    /// decision model.
+    async fn default_decision_model(
+        &self,
+        requested: Option<&str>,
+    ) -> Result<Arc<super::decision_model::LoadedDecisionModel>, InferenceError> {
+        let models = self.models.read().await;
+        if let Some(id) = requested
+            && let Some(handle) = models.get(id)
+        {
+            return match handle {
+                LoadedHandle::Decision(m) => Ok(Arc::clone(m)),
+                _ => Err(InferenceError::WrongModality {
+                    model_id: id.to_string(),
+                }),
+            };
+        }
+        let mut decision: Vec<&Arc<super::decision_model::LoadedDecisionModel>> = models
+            .values()
+            .filter_map(|h| match h {
+                LoadedHandle::Decision(m) => Some(m),
+                _ => None,
+            })
+            .collect();
+        decision.sort_by(|a, b| a.model_id.cmp(&b.model_id));
+        decision
+            .first()
+            .map(|m| Arc::clone(m))
+            .ok_or_else(|| InferenceError::ModelNotLoaded("no decision model is loaded".into()))
+    }
+
     /// Run one text-to-image generation (#198). This is the internal
     /// path the `/v1/images/generations` handler (#200) fronts;
     /// admission, validation, tokenization, and PNG encoding happen
@@ -5322,6 +5685,9 @@ impl CandleHarness {
                 // read them; populated at the prefill→decode boundary inside.
                 let mut prefill_ms_measured: u32 = 0;
                 let mut decode_start: Option<std::time::Instant> = None;
+                let mut phase_forward = std::time::Duration::ZERO;
+                let mut phase_sample = std::time::Duration::ZERO;
+                let mut phase_emit = std::time::Duration::ZERO;
 
                 // Hoisted out of `'work` so the completion path below
                 // can report it (#269) — the reuse count is decided
@@ -5482,6 +5848,12 @@ impl CandleHarness {
                         };
                     // Decode-phase timer for the Finish prefill/decode split (#85).
                     decode_start = Some(std::time::Instant::now());
+                    // Phase split (see `PhaseTiming`). Instrumented here
+                    // as well as on the single-GPU path so a healthy
+                    // model is a control for a slow one: both archs share
+                    // this loop, this sampler and a 248,320 vocabulary,
+                    // so a phase that is cheap here and expensive there
+                    // is a property of the arch, not of the loop.
 
                     if Some(next_token) == eos_id {
                         finish_reason = FinishReason::Stop;
@@ -5568,6 +5940,7 @@ impl CandleHarness {
                         }
 
                         for index in 0..max_new.saturating_sub(1) {
+                            let t_forward = std::time::Instant::now();
                             let logits_vec = match pool
                                 .generate_step(
                                     &model_id,
@@ -5583,6 +5956,8 @@ impl CandleHarness {
                                     break 'work;
                                 }
                             };
+                            phase_forward += t_forward.elapsed();
+                            let t_sample = std::time::Instant::now();
                             let logits = match Tensor::new(logits_vec.as_slice(), &Device::Cpu) {
                                 Ok(t) => t,
                                 Err(e) => {
@@ -5611,6 +5986,11 @@ impl CandleHarness {
                                     break 'work;
                                 }
                             };
+                            phase_sample += t_sample.elapsed();
+                            // NB: this per-step VRAM query is a second
+                            // channel round-trip per token. It is charged
+                            // to neither phase, so it lands in the
+                            // residual between the three and `decode_ms`.
                             // Always await the query (even when the
                             // trace! is filtered out by RUST_LOG): the
                             // channel hop is ~tens of µs, comparable to
@@ -5630,6 +6010,7 @@ impl CandleHarness {
                                 finish_reason = FinishReason::Stop;
                                 break;
                             }
+                            let t_emit = std::time::Instant::now();
                             all_tokens.push(next_token);
                             match handle_tool_call_marker(
                                 next_token,
@@ -5711,6 +6092,7 @@ impl CandleHarness {
                                     "TP stream: decode_stream step failed"
                                 ),
                             }
+                            phase_emit += t_emit.elapsed();
                         }
                     }
                 }
@@ -5775,8 +6157,24 @@ impl CandleHarness {
                         decode_ms = decode_start
                             .map(|d| d.elapsed().as_millis() as u32)
                             .unwrap_or(0),
+                        forward_ms = phase_forward.as_millis(),
+                        sample_ms = phase_sample.as_millis(),
+                        emit_ms = phase_emit.as_millis(),
                         "chat_completion (stream): done"
                     );
+                    crate::metrics::record_finish(
+                        &model_id,
+                        prefill_ms_measured,
+                        decode_start
+                            .map(|d| d.elapsed().as_millis() as u32)
+                            .unwrap_or(0),
+                        all_tokens.len() as u32,
+                        Some((
+                            phase_forward.as_millis() as u32,
+                            phase_sample.as_millis() as u32,
+                            phase_emit.as_millis() as u32,
+                        )),
+                    );
                     let _ = tx
                         .send(InferenceEvent::Finish {
                             reason: finish_reason,
@@ -5790,6 +6188,11 @@ impl CandleHarness {
                                     .map(|d| d.elapsed().as_millis() as u32)
                                     .unwrap_or(0),
                                 prefill_tokens: prompt_len as u32,
+                                phases: Some(PhaseTiming {
+                                    forward_ms: phase_forward.as_millis() as u32,
+                                    sample_ms: phase_sample.as_millis() as u32,
+                                    emit_ms: phase_emit.as_millis() as u32,
+                                }),
                             }),
                         })
                         .await;
@@ -6219,6 +6622,33 @@ pub enum InferenceError {
     Other(#[from] anyhow::Error),
 }
 
+/// Why a `/v1/systemone` request was not answered (#338): a request the
+/// reference server would also reject (its status and `detail` text), or
+/// a serving failure shared with the other endpoints.
+#[derive(Debug)]
+pub enum SystemOneError {
+    Request(super::decision::DecisionRequestError),
+    Inference(InferenceError),
+}
+
+impl From<super::decision::DecisionRequestError> for SystemOneError {
+    fn from(e: super::decision::DecisionRequestError) -> Self {
+        SystemOneError::Request(e)
+    }
+}
+
+impl From<InferenceError> for SystemOneError {
+    fn from(e: InferenceError) -> Self {
+        SystemOneError::Inference(e)
+    }
+}
+
+impl From<super::admission::AdmissionRejection> for SystemOneError {
+    fn from(e: super::admission::AdmissionRejection) -> Self {
+        SystemOneError::Inference(e.into())
+    }
+}
+
 impl From<super::admission::AdmissionRejection> for InferenceError {
     fn from(rejection: super::admission::AdmissionRejection) -> Self {
         use super::admission::AdmissionRejection;
@@ -7236,6 +7666,9 @@ async fn run_inference_via_worker(
 #[cfg(feature = "cuda")]
 #[allow(clippy::too_many_arguments)]
 async fn stream_inference_via_worker(
+    // Owned: this runs in a spawned task, and the only other source of
+    // the model name here is the tracing span, which is not a value.
+    model_id: String,
     worker: Arc<super::device_worker::DeviceWorkerHandle>,
     handle: super::device_worker::ArchHandle,
     tokenizer: Tokenizer,
@@ -7369,6 +7802,13 @@ async fn stream_inference_via_worker(
     };
     // Decode-phase timer for the Finish prefill/decode split (#85).
     let decode_start = std::time::Instant::now();
+    // Phase split of the decode loop. Each boundary below is already a
+    // synchronisation point -- the worker hands back CPU-side logits --
+    // so this adds no device sync and measures elapsed work rather than
+    // kernel-launch time.
+    let mut phase_forward = std::time::Duration::ZERO;
+    let mut phase_sample = std::time::Duration::ZERO;
+    let mut phase_emit = std::time::Duration::ZERO;
 
     // Per-token routing. `tokenizers::DecodeStream` carries five
     // generic parameters (`M, N, PT, PP, D`) which makes naming
@@ -7475,10 +7915,14 @@ async fn stream_inference_via_worker(
     }
 
     for index in 0..max_new.saturating_sub(1) {
+        let t_forward = std::time::Instant::now();
         let logits_vec = worker
             .forward_logits(handle, vec![next_token], prompt_len + index)
             .await
             .map_err(|e| anyhow::anyhow!("decode step {index}: {e:#}"))?;
+        phase_forward += t_forward.elapsed();
+
+        let t_sample = std::time::Instant::now();
         let logits = Tensor::new(logits_vec.as_slice(), &Device::Cpu)?;
         next_token = match sample_with_penalty(
             &logits,
@@ -7498,11 +7942,15 @@ async fn stream_inference_via_worker(
                 return Err(e);
             }
         };
+        phase_sample += t_sample.elapsed();
         if Some(next_token) == eos_id {
             finish_reason = FinishReason::Stop;
             break;
         }
-        if !route_token!(next_token) {
+        let t_emit = std::time::Instant::now();
+        let routed = route_token!(next_token);
+        phase_emit += t_emit.elapsed();
+        if !routed {
             return Ok(finish_reason.as_openai_str().to_string());
         }
     }
@@ -7524,8 +7972,22 @@ async fn stream_inference_via_worker(
         finish_reason = ?finish_reason,
         prefill_ms = prefill_elapsed.as_millis(),
         decode_ms = decode_start.elapsed().as_millis(),
+        forward_ms = phase_forward.as_millis(),
+        sample_ms = phase_sample.as_millis(),
+        emit_ms = phase_emit.as_millis(),
         "chat_completion (stream): done"
     );
+    crate::metrics::record_finish(
+        &model_id,
+        prefill_elapsed.as_millis() as u32,
+        decode_start.elapsed().as_millis() as u32,
+        all_tokens.len() as u32,
+        Some((
+            phase_forward.as_millis() as u32,
+            phase_sample.as_millis() as u32,
+            phase_emit.as_millis() as u32,
+        )),
+    );
     let _ = tx
         .send(InferenceEvent::Finish {
             reason: finish_reason,
@@ -7537,6 +7999,11 @@ async fn stream_inference_via_worker(
                 prefill_ms: prefill_elapsed.as_millis() as u32,
                 decode_ms: decode_start.elapsed().as_millis() as u32,
                 prefill_tokens: prefill_prompt_len as u32,
+                phases: Some(PhaseTiming {
+                    forward_ms: phase_forward.as_millis() as u32,
+                    sample_ms: phase_sample.as_millis() as u32,
+                    emit_ms: phase_emit.as_millis() as u32,
+                }),
             }),
         })
         .await;
@@ -7757,6 +8224,7 @@ fn run_inference_streaming(
             prefill_ms: prefill_elapsed.as_millis() as u32,
             decode_ms: decode_start.elapsed().as_millis() as u32,
             prefill_tokens: prompt_tokens.len() as u32,
+            phases: None,
         }),
     });
     Ok(())

crates/neuron/src/harness/capability.rs

@@ -57,6 +57,9 @@ pub fn classify_files(
 ) -> Vec<String> {
     match preflight::classify(filenames) {
         SourceFormat::Diffusers => vec!["image".to_string()],
+        // A decision model answers typed questions (`/v1/systemone`);
+        // it generates no text, so it must not claim "text".
+        SourceFormat::DecisionModel => vec!["decision".to_string()],
         // A repo with no recognised weights (tokenizer-only, or an empty
         // entry) is not servable, so claiming "text" would be a lie of
         // the same kind this module exists to remove.
@@ -101,6 +104,22 @@ mod tests {
         );
     }
 
+    #[test]
+    fn laya_repo_is_a_decision_model() {
+        let files = [
+            "rl_agent_config.json",
+            "encoder/config.json",
+            "model.safetensors",
+            "tokenizer/tokenizer.json",
+        ];
+        assert_eq!(
+            classify_files(&files, || panic!(
+                "config must not be read for a decision model"
+            )),
+            vec!["decision".to_string()]
+        );
+    }
+
     #[test]
     fn dense_repo_without_vision_config_is_text_only() {
         let files = ["config.json", "tokenizer.json", "model.safetensors"];

crates/neuron/src/harness/decision/config.rs

@@ -0,0 +1,233 @@
+//! Per-checkpoint decision settings, read from `rl_agent_config.json`.
+//!
+//! Parsed with [`Json`], not `serde_json`: the fitted temperatures are
+//! written with 17 significant digits, and `serde_json`'s default float
+//! parser is not correctly rounded (it reads `1.7601518630981445` as
+//! `…443`).
+
+use cortex_core::decisions::Json;
+use std::collections::HashMap;
+
+/// Default token budgets when a checkpoint's config omits them.
+pub const DEFAULT_MAX_LEN: usize = 512;
+pub const DEFAULT_HEAD_MAX_LEN: usize = 192;
+
+/// Bounds on a usable calibration temperature. A fitted value below 1
+/// sharpens instead of softening; the shipped `choice:11+` bucket is 0.1006,
+/// which would publish a coin flip as a certainty, so the reference refuses
+/// to apply anything outside this range.
+pub const TEMP_MIN: f64 = 0.5;
+pub const TEMP_MAX: f64 = 5.0;
+
+/// A temperature confined to [`TEMP_MIN`, `TEMP_MAX`]; a non-finite or
+/// non-numeric value falls back to 1.0.
+pub fn clamp_temperature(t: Option<f64>) -> f64 {
+    match t {
+        Some(t) if t.is_finite() => t.clamp(TEMP_MIN, TEMP_MAX),
+        _ => 1.0,
+    }
+}
+
+/// Python's `float(v)` for a JSON value: numbers, bools, and numeric strings.
+fn py_float(v: &Json) -> Option<f64> {
+    match v {
+        Json::Int(d) => d.parse().ok(),
+        Json::Float(x) => Some(*x),
+        Json::Bool(b) => Some(if *b { 1.0 } else { 0.0 }),
+        Json::String(s) => s.trim().parse().ok(),
+        _ => None,
+    }
+}
+
+/// One set of calibration temperatures: per question type, overridden per
+/// `(type, option-count bucket)` (see [`temp_bucket`]). Already clamped.
+#[derive(Debug, Clone, PartialEq)]
+pub struct Temperatures {
+    /// Indexed by [`super::QType`] (`choice`, `score`, `noul`).
+    pub by_type: [f64; 3],
+    /// `"choice:3-5"` → temperature, and so on.
+    pub by_options: HashMap<String, f64>,
+}
+
+impl Temperatures {
+    /// The temperature for a question of type index `qtype` with `k` options.
+    pub fn for_question(&self, qtype: usize, k: usize) -> f64 {
+        self.by_options
+            .get(&temp_bucket(qtype, k))
+            .copied()
+            .unwrap_or(self.by_type[qtype])
+    }
+
+    fn from_json(
+        temperature: Option<&Json>,
+        by_options: Option<&Json>,
+        fallback: Option<[f64; 3]>,
+    ) -> Result<Self, String> {
+        let raw = match temperature {
+            None => fallback.unwrap_or([1.0; 3]).map(Some),
+            Some(Json::Array(a)) if a.len() == 3 => {
+                [py_float(&a[0]), py_float(&a[1]), py_float(&a[2])]
+            }
+            Some(other) => {
+                return Err(format!(
+                    "temperature must be a list of 3 floats, got {}",
+                    other.py_repr()
+                ));
+            }
+        };
+        let by_options = match by_options {
+            Some(Json::Object(m)) => m
+                .iter()
+                .map(|(k, v)| (k.clone(), clamp_temperature(py_float(v))))
+                .collect(),
+            _ => HashMap::new(),
+        };
+        Ok(Self {
+            by_type: raw.map(clamp_temperature),
+            by_options,
+        })
+    }
+}
+
+/// The calibration bucket of a question: its type and option count.
+pub fn temp_bucket(qtype: usize, k: usize) -> String {
+    let size = match k {
+        0..=2 => "2",
+        3..=5 => "3-5",
+        6..=10 => "6-10",
+        _ => "11+",
+    };
+    format!("{}:{size}", super::QType::NAMES[qtype])
+}
+
+/// What the sequence builder and the calibration need from a checkpoint.
+#[derive(Debug, Clone, PartialEq)]
+pub struct DecisionConfig {
+    /// Total sequence budget.
+    pub max_len: usize,
+    /// Budget for the question text and every option together.
+    pub head_max_len: usize,
+    pub temperatures: Temperatures,
+    /// Per-language overrides, keyed by the lower-cased primary subtag
+    /// (`"de"` for `de-AT`). Empty for the shipped checkpoints; an operator
+    /// may fit them per language.
+    pub lang_temperatures: HashMap<String, Temperatures>,
+}
+
+impl DecisionConfig {
+    /// Parse a checkpoint's `rl_agent_config.json`.
+    pub fn from_json(text: &str) -> Result<Self, String> {
+        let v = Json::parse(text.as_bytes()).map_err(|e| format!("rl_agent_config.json: {e}"))?;
+        let usize_or = |key: &str, default: usize| match v.get(key) {
+            Some(Json::Int(d)) => d.parse().unwrap_or(default),
+            _ => default,
+        };
+        Ok(Self {
+            max_len: usize_or("max_len", DEFAULT_MAX_LEN),
+            head_max_len: usize_or("head_max_len", DEFAULT_HEAD_MAX_LEN),
+            temperatures: Temperatures::from_json(
+                v.get("temperature"),
+                v.get("temperature_by_options"),
+                None,
+            )?,
+            lang_temperatures: HashMap::new(),
+        })
+    }
+
+    /// Add per-language temperatures, the shape the reference takes them in:
+    /// `{"de": {"temperature": [..3], "temperature_by_options": {..}}}`. A
+    /// language that omits `temperature` inherits the checkpoint's raw
+    /// per-type values.
+    pub fn with_lang_temperatures(mut self, spec: &Json) -> Result<Self, String> {
+        let Json::Object(m) = spec else {
+            return Err("lang_temperatures must be an object".into());
+        };
+        for (lang, cfg) in m {
+            let t = Temperatures::from_json(
+                cfg.get("temperature"),
+                cfg.get("temperature_by_options"),
+                Some(self.temperatures.by_type),
+            )
+            .map_err(|e| format!("language override {lang:?}: {e}"))?;
+            let primary = lang.split('-').next().unwrap_or(lang).to_lowercase();
+            self.lang_temperatures.insert(primary, t);
+        }
+        Ok(self)
+    }
+
+    /// The temperatures to apply for a request detected (or declared) as
+    /// `lang`, which may carry a region (`de-AT`).
+    pub fn temperatures_for(&self, lang: Option<&str>) -> &Temperatures {
+        lang.and_then(|l| {
+            let primary = l.split('-').next().unwrap_or(l).to_lowercase();
+            self.lang_temperatures.get(&primary)
+        })
+        .unwrap_or(&self.temperatures)
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    const ENGLISH: &str = r#"{"encoder": "answerdotai/ModernBERT-large", "head_layers": 2,
+        "max_len": 512, "head_max_len": 192,
+        "temperature": [1.6369030475616455, 1.2514300346374512, 1.983399510383606],
+        "temperature_by_options": {"choice:3-5": 1.7601518630981445, "choice:11+": 0.10058280825614929}}"#;
+
+    #[test]
+    fn shipped_choice_11_bucket_is_clamped() {
+        let c = DecisionConfig::from_json(ENGLISH).unwrap();
+        assert_eq!(c.temperatures.for_question(0, 12), TEMP_MIN);
+        assert_eq!(c.temperatures.for_question(0, 4), 1.7601518630981445);
+        // No bucket: the per-type value.
+        assert_eq!(c.temperatures.for_question(1, 7), 1.2514300346374512);
+    }
+
+    #[test]
+    fn buckets_split_where_the_reference_splits() {
+        let b: Vec<_> = [1, 2, 3, 5, 6, 10, 11]
+            .iter()
+            .map(|&k| temp_bucket(2, k))
+            .collect();
+        assert_eq!(
+            b,
+            [
+                "noul:2",
+                "noul:2",
+                "noul:3-5",
+                "noul:3-5",
+                "noul:6-10",
+                "noul:6-10",
+                "noul:11+"
+            ]
+        );
+    }
+
+    #[test]
+    fn language_overrides_match_on_primary_subtag() {
+        let c = DecisionConfig::from_json(ENGLISH)
+            .unwrap()
+            .with_lang_temperatures(
+                &Json::parse(br#"{"DE": {"temperature": [1.0, 1.4, 9.0]}}"#).unwrap(),
+            )
+            .unwrap();
+        assert_eq!(
+            c.temperatures_for(Some("de-AT")).by_type,
+            [1.0, 1.4, TEMP_MAX]
+        );
+        assert_eq!(c.temperatures_for(Some("fr")), &c.temperatures);
+        assert_eq!(c.temperatures_for(None), &c.temperatures);
+    }
+
+    #[test]
+    fn bad_temperatures_are_refused_or_neutralised() {
+        assert!(DecisionConfig::from_json(r#"{"temperature": [1.0]}"#).is_err());
+        let c = DecisionConfig::from_json(r#"{"temperature": [" 2.5 ", "x", null]}"#).unwrap();
+        assert_eq!(c.temperatures.by_type, [2.5, 1.0, 1.0]);
+        assert_eq!(
+            (c.max_len, c.head_max_len),
+            (DEFAULT_MAX_LEN, DEFAULT_HEAD_MAX_LEN)
+        );
+    }
+}

crates/neuron/src/harness/decision/handler_tests.rs

@@ -0,0 +1,249 @@
+//! `CandleHarness::systemone` end to end: request in, routed, validated,
+//! scored and shaped response out (#338).
+//!
+//! The tiny checkpoint (`testdata/laya/tiny/`) exercises the plumbing in
+//! CI. The released-weights test replays every reference case through
+//! the same entry point the HTTP handler calls and compares the whole
+//! response — answers, usage and routing — with what the reference
+//! server returned; it needs `LAYA_REFERENCE_DIR`.
+
+use std::path::{Path, PathBuf};
+use std::sync::Arc;
+use std::sync::atomic::AtomicBool;
+
+use cortex_core::decisions::{Json, SystemOneRequest, limits};
+
+use crate::harness::candle::{CandleHarness, SystemOneError};
+use crate::harness::decision_model::{CheckpointFiles, DecisionCheckpoint, LoadedDecisionModel};
+
+fn testdata() -> PathBuf {
+    Path::new(env!("CARGO_MANIFEST_DIR")).join("src/harness/testdata/laya")
+}
+
+fn harness() -> Arc<CandleHarness> {
+    CandleHarness::new(
+        "http://localhost:13131".into(),
+        &crate::config::CandleHarnessConfig::default(),
+    )
+}
+
+/// A decision model made of the given `(routing name, directory)`
+/// checkpoints, loaded on the CPU in f32.
+async fn decision_model(model_id: &str, checkpoints: &[(&str, PathBuf)]) -> LoadedDecisionModel {
+    let mut loaded = Vec::new();
+    for (name, dir) in checkpoints {
+        let files = CheckpointFiles {
+            tokenizer: if dir.join("tokenizer/tokenizer.json").exists() {
+                dir.join("tokenizer/tokenizer.json")
+            } else {
+                dir.join("tokenizer.json")
+            },
+            dir: dir.clone(),
+        };
+        loaded.push(
+            DecisionCheckpoint::load(name.to_string(), files, model_id, None)
+                .await
+                .unwrap_or_else(|e| panic!("load {name}: {e:#}")),
+        );
+    }
+    LoadedDecisionModel {
+        model_id: model_id.to_string(),
+        spec: serde_json::from_value(serde_json::json!({
+            "model_id": model_id,
+            "harness": "candle",
+        }))
+        .expect("model spec"),
+        devices: vec![0],
+        checkpoints: loaded,
+        poisoned: Arc::new(AtomicBool::new(false)),
+        admission: crate::harness::admission::AdmissionController::new(
+            &crate::config::AdmissionConfig::default(),
+        ),
+    }
+}
+
+fn request(body: &str) -> SystemOneRequest {
+    SystemOneRequest::parse(body.as_bytes(), limits::DEFAULT_MAX_TOKEN_BUDGET)
+        .expect("valid request")
+}
+
+async fn tiny_harness() -> Arc<CandleHarness> {
+    let h = harness();
+    h.register_decision_model(
+        decision_model("test/tiny", &[("english", testdata().join("tiny"))]).await,
+    )
+    .await;
+    h
+}
+
+#[tokio::test]
+async fn answers_every_question_in_order_with_exact_usage() {
+    let h = tiny_harness().await;
+    let req = request(
+        r#"{"state": "a b a", "questions": {
+            "z_last_alphabetically": {"type": "noul", "instructions": "a?"},
+            "a_first": {"type": "choice", "instructions": "b?", "criteria": ["a", "b", "a b"]}
+        }}"#,
+    );
+    let resp = h.systemone(None, req, None).await.expect("answered");
+    let json = serde_json::to_value(&resp).unwrap();
+    // Answers come back in request order, not sorted.
+    let keys: Vec<&str> = resp.answers.0.iter().map(|(k, _)| k.as_str()).collect();
+    assert_eq!(keys, ["z_last_alphabetically", "a_first"]);
+    assert_eq!(json["answers"]["z_last_alphabetically"]["type"], "noul");
+    let probs = json["answers"]["a_first"]["probabilities"]
+        .as_object()
+        .unwrap();
+    let total: f64 = probs.values().map(|p| p.as_f64().unwrap()).sum();
+    assert!((total - 1.0).abs() < 1e-3, "{probs:?}");
+    // Usage is the exact number of tokens encoded, output tokens zero.
+    assert!(json["usage"]["input_tokens"].as_u64().unwrap() > 0);
+    assert_eq!(json["usage"]["output_tokens"], 0);
+    assert_eq!(json["routing"]["model"], "english");
+    assert_eq!(json["routing"]["repo"], "test/tiny");
+}
+
+#[tokio::test]
+async fn pinning_a_checkpoint_the_model_lacks_is_404() {
+    let h = tiny_harness().await;
+    let req = request(
+        r#"{"state": "a", "model": "multilingual",
+            "questions": {"q": {"type": "noul", "instructions": "a?"}}}"#,
+    );
+    match h.systemone(None, req, None).await {
+        Err(SystemOneError::Request(e)) => assert_eq!(e.status, 404, "{}", e.detail),
+        other => panic!("expected a 404, got {other:?}"),
+    }
+}
+
+#[tokio::test]
+async fn routing_to_a_missing_checkpoint_falls_back_and_says_so() {
+    // Devanagari routes to `multilingual`, which this model lacks.
+    let h = tiny_harness().await;
+    let req = request(
+        r#"{"state": "मुझे पैसे वापस चाहिए",
+            "questions": {"q": {"type": "noul", "instructions": "a?"}}}"#,
+    );
+    let resp = h.systemone(None, req, None).await.expect("answered");
+    let routing = resp.routing.expect("routing");
+    assert_eq!(routing.model, "english");
+    assert!(
+        routing
+            .reason
+            .contains("multilingual checkpoint is not loaded"),
+        "{}",
+        routing.reason
+    );
+}
+
+#[tokio::test]
+async fn invalid_questions_are_rejected_before_admission() {
+    let h = tiny_harness().await;
+    let req =
+        request(r#"{"state": "a", "questions": {"q": {"type": "maybe", "instructions": "a?"}}}"#);
+    match h.systemone(None, req, None).await {
+        Err(SystemOneError::Request(e)) => assert_eq!(e.status, 422, "{}", e.detail),
+        other => panic!("expected a 422, got {other:?}"),
+    }
+}
+
+#[tokio::test]
+async fn a_gateway_routed_model_that_is_not_a_decision_model_is_refused() {
+    let h = tiny_harness().await;
+    let req = request(r#"{"state": "a", "questions": {}}"#);
+    match h.systemone(Some("not/loaded"), req, None).await {
+        Err(SystemOneError::Inference(crate::harness::candle::InferenceError::ModelNotLoaded(
+            id,
+        ))) => {
+            assert_eq!(id, "not/loaded")
+        }
+        other => panic!("expected ModelNotLoaded, got {other:?}"),
+    }
+}
+
+/// Every reference case through `systemone`, compared with the reference
+/// server's whole response. Routing and usage must match exactly;
+/// probabilities to within the engine's measured f32 distance from the
+/// reference (a last-digit wobble after 4-decimal rounding); the chosen
+/// answer exactly.
+#[tokio::test]
+#[ignore = "needs the released weights: set LAYA_REFERENCE_DIR"]
+async fn released_family_answers_like_the_reference_server() {
+    let root = PathBuf::from(std::env::var("LAYA_REFERENCE_DIR").expect("LAYA_REFERENCE_DIR"));
+    let h = harness();
+    h.register_decision_model(
+        decision_model(
+            "convaiinnovations/laya",
+            &[
+                ("english", root.clone()),
+                ("multilingual", root.join("multilingual")),
+                ("typed-decisions", root.join("typed-decisions")),
+            ],
+        )
+        .await,
+    )
+    .await;
+
+    let text = std::fs::read_to_string(testdata().join("reference.json")).unwrap();
+    let reference = Json::parse(text.as_bytes()).unwrap();
+    let Some(Json::Array(cases)) = reference.get("cases") else {
+        panic!("no cases")
+    };
+    let mut max_dp = 0.0f64;
+    for case in cases {
+        let name = case.get("name").and_then(Json::as_str).unwrap();
+        let body = serde_json::to_vec(case.get("request").unwrap()).unwrap();
+        let req = SystemOneRequest::parse(&body, limits::DEFAULT_MAX_TOKEN_BUDGET).unwrap();
+        let resp = h
+            .systemone(Some("convaiinnovations/laya"), req, None)
+            .await
+            .unwrap_or_else(|e| panic!("{name}: {e:?}"));
+        let want_json = case.get("response").unwrap();
+        // Routing and usage compared as serialized text through `Json`:
+        // exact, key order included. serde_json would re-round the
+        // reference's floats on the way in.
+        for key in ["routing", "usage", "model"] {
+            let got = match key {
+                "routing" => serde_json::to_string(&resp.routing).unwrap(),
+                "usage" => serde_json::to_string(&resp.usage).unwrap(),
+                _ => serde_json::to_string(&resp.model).unwrap(),
+            };
+            let want = serde_json::to_string(want_json.get(key).unwrap()).unwrap();
+            assert_eq!(got, want, "{name}: {key}");
+        }
+        let got = serde_json::to_value(&resp).unwrap();
+        let want: serde_json::Value =
+            serde_json::from_str(&serde_json::to_string(want_json).unwrap()).unwrap();
+        let (ga, wa) = (
+            got["answers"].as_object().unwrap(),
+            want["answers"].as_object().unwrap(),
+        );
+        assert_eq!(
+            ga.keys().collect::<Vec<_>>(),
+            wa.keys().collect::<Vec<_>>(),
+            "{name}: answer ids"
+        );
+        for (qid, w) in wa {
+            let g = &ga[qid];
+            for key in ["choice", "type", "legend"] {
+                assert_eq!(g.get(key), w.get(key), "{name}/{qid}: {key}");
+            }
+            if let Some(wp) = w["probabilities"].as_object() {
+                for (opt, p) in wp {
+                    let d = (g["probabilities"][opt].as_f64().unwrap() - p.as_f64().unwrap()).abs();
+                    max_dp = max_dp.max(d);
+                }
+            }
+            for key in ["noul", "score", "confidence", "answer_confidence"] {
+                if let (Some(gv), Some(wv)) = (
+                    g.get(key).and_then(|v| v.as_f64()),
+                    w.get(key).and_then(|v| v.as_f64()),
+                ) {
+                    max_dp = max_dp.max((gv - wv).abs());
+                }
+            }
+        }
+    }
+    eprintln!("released family: max |Δ| over probabilities and confidences {max_dp:.1e}");
+    assert!(max_dp <= 5e-4, "max |Δ| {max_dp:e}");
+}

crates/neuron/src/harness/decision/mod.rs

@@ -0,0 +1,59 @@
+//! Laya decision models: everything around the forward pass (#335).
+//!
+//! A Laya checkpoint answers typed questions over a *state* (text, a JSON
+//! document, a conversation) in one encoder forward pass: each question
+//! becomes one sequence,
+//!
+//! ```text
+//! [CLS] "<type> question: <instructions>" [SEP] [MASK] opt0 [MASK] opt1 … [SEP] state [SEP]
+//! ```
+//!
+//! and the decision head scores the hidden state at each option's `[MASK]`
+//! marker. This module is the part that touches no tensor:
+//!
+//! - [`question`] — validate and normalise the question definitions, and
+//!   render each option's text;
+//! - [`sequence`] — assemble the token sequences under the model's token
+//!   budgets ([`prepare`]);
+//! - [`shape`] — turn per-option logits into calibrated, typed answers
+//!   ([`shape()`]);
+//! - [`config`] / [`tokenizer`] — the per-checkpoint pieces those need.
+//!
+//! The reference is the `laya` Python SDK (`laya/common.py`,
+//! `laya/agent.py`, `laya/serve.py`), whose behaviour this reproduces token
+//! for token: the parity tests replay fixtures recorded from it
+//! (`script/laya-reference.py`). The calibration arithmetic follows numpy's
+//! float32 evaluation order, so rounded probabilities match to the last
+//! published digit.
+//!
+//! The wire types (`/v1/systemone`) live in [`cortex_core::decisions`].
+
+pub mod config;
+pub mod question;
+pub mod route;
+pub mod sequence;
+pub mod shape;
+pub mod tokenizer;
+
+#[cfg(test)]
+mod handler_tests;
+
+pub use config::DecisionConfig;
+pub use cortex_core::decisions::DecisionRequestError;
+pub use question::{QType, Question};
+pub use sequence::{Budget, OptionStats, Prepared, Row, prepare};
+pub use shape::{RowOutput, shape};
+pub use tokenizer::{DecisionTokenizer, SpecialTokens, TextEncoder};
+
+/// The constant `model` field of every response: the decision head's name.
+/// Which checkpoint answered is reported in `routing`.
+pub const DECISION_MODEL_NAME: &str = "laya-rl-agent";
+
+/// A request error in the reference's `ValueError` class: the server
+/// answers it with 422 and the message as `detail`.
+pub(crate) fn unprocessable(detail: impl Into<String>) -> DecisionRequestError {
+    DecisionRequestError::new(422, detail)
+}
+
+#[cfg(test)]
+mod parity_tests;

crates/neuron/src/harness/decision/parity_tests.rs

@@ -0,0 +1,209 @@
+//! Replays the reference fixtures (`testdata/laya/reference.json`, recorded
+//! by `script/laya-reference.py`) through [`prepare`] and [`shape`].
+//!
+//! Token ids and marker positions must match exactly. The answers and usage
+//! must serialize to exactly the same JSON — same keys, same order, same
+//! rounded values — when fed the reference's own raw logits.
+//!
+//! The English tokenizer is committed. The multilingual one is 34 MB, so the
+//! multilingual cases run only when `LAYA_ML_TOKENIZER_DIR` names a copy of
+//! the checkpoint's `multilingual/tokenizer/` directory.
+
+use super::*;
+use cortex_core::decisions::{Json, SystemOneRequest, limits};
+use std::path::{Path, PathBuf};
+
+fn testdata() -> PathBuf {
+    Path::new(env!("CARGO_MANIFEST_DIR")).join("src/harness/testdata/laya")
+}
+
+fn reference() -> Json {
+    Json::parse(&std::fs::read(testdata().join("reference.json")).expect("fixtures")).expect("json")
+}
+
+fn to_value(j: &Json) -> serde_json::Value {
+    serde_json::to_value(j).expect("serializable")
+}
+
+fn tokenizer_for(checkpoint: &str) -> Option<DecisionTokenizer> {
+    let dir = match checkpoint {
+        // typed-decisions ships the English tokenizer, byte for byte.
+        "english" | "typed-decisions" => testdata().join("tokenizer"),
+        "multilingual" => PathBuf::from(std::env::var_os("LAYA_ML_TOKENIZER_DIR")?),
+        other => panic!("unknown checkpoint {other}"),
+    };
+    Some(DecisionTokenizer::from_dir(&dir).expect("tokenizer"))
+}
+
+fn config_for(reference: &Json, checkpoint: &str) -> DecisionConfig {
+    let cfg = reference
+        .get("header")
+        .and_then(|h| h.get("checkpoints"))
+        .and_then(|c| c.get(checkpoint))
+        .and_then(|c| c.get("cfg"))
+        .expect("checkpoint cfg");
+    DecisionConfig::from_json(&serde_json::to_string(cfg).unwrap()).unwrap()
+}
+
+fn array(j: &Json) -> &[Json] {
+    match j {
+        Json::Array(a) => a,
+        other => panic!("expected a list, got {other:?}"),
+    }
+}
+
+fn num(j: &Json) -> f64 {
+    match j {
+        Json::Int(d) => d.parse().unwrap(),
+        Json::Float(x) => *x,
+        other => panic!("expected a number, got {other:?}"),
+    }
+}
+
+/// Replay every case whose tokenizer is available; return how many ran per
+/// checkpoint.
+fn replay(filter: impl Fn(&str) -> bool) -> Vec<(String, usize)> {
+    let reference = reference();
+    let mut ran: Vec<(String, usize)> = Vec::new();
+    for case in array(reference.get("cases").unwrap()) {
+        let name = case.get("name").and_then(Json::as_str).unwrap();
+        if !filter(name) {
+            continue;
+        }
+        let response = case.get("response").unwrap();
+        let routing = response.get("routing").unwrap();
+        let checkpoint = routing.get("model").and_then(Json::as_str).unwrap();
+        let Some(tok) = tokenizer_for(checkpoint) else {
+            continue;
+        };
+        let cfg = config_for(&reference, checkpoint);
+
+        let body = serde_json::to_vec(case.get("request").unwrap()).unwrap();
+        let req = SystemOneRequest::parse(&body, limits::DEFAULT_MAX_TOKEN_BUDGET)
+            .unwrap_or_else(|e| panic!("{name}: {e}"));
+        let budget = Budget {
+            max_len: req.max_len,
+            head_max_len: req.head_max_len,
+        };
+        let prepared = prepare(&req.state, &req.questions, &cfg, &tok, budget)
+            .unwrap_or_else(|e| panic!("{name}: {e}"));
+
+        let items = array(case.get("items").unwrap());
+        assert_eq!(prepared.rows.len(), items.len(), "{name}: row count");
+        let mut outputs = Vec::new();
+        for (i, (row, item)) in prepared.rows.iter().zip(items).enumerate() {
+            let ids: Vec<u32> = array(item.get("input_ids").unwrap())
+                .iter()
+                .map(|v| num(v) as u32)
+                .collect();
+            let markers: Vec<usize> = array(item.get("markers").unwrap())
+                .iter()
+                .map(|v| num(v) as usize)
+                .collect();
+            assert_eq!(row.input_ids, ids, "{name}: row {i} input_ids");
+            assert_eq!(row.markers, markers, "{name}: row {i} markers");
+            assert_eq!(row.qtype as usize, num(item.get("qtype").unwrap()) as usize);
+            let opts = item.get("options").unwrap();
+            let s = prepared.stats[i];
+            assert_eq!(
+                (s.options, s.options_distinct, s.tokens_per_option),
+                (
+                    num(opts.get("options").unwrap()) as usize,
+                    num(opts.get("options_distinct").unwrap()) as usize,
+                    match opts.get("tokens_per_option").unwrap() {
+                        Json::Null => None,
+                        v => Some(num(v) as usize),
+                    }
+                ),
+                "{name}: row {i} option stats"
+            );
+            outputs.push(RowOutput {
+                logits: array(item.get("logits").unwrap())
+                    .iter()
+                    .map(|v| num(v) as f32)
+                    .collect(),
+                act_probability: num(&array(item.get("act").unwrap())[0]) as f32,
+            });
+        }
+
+        let lang = routing
+            .get("detection")
+            .and_then(|d| d.get("language"))
+            .and_then(Json::as_str);
+        let (answers, usage) = shape(&prepared, &outputs, &cfg, lang);
+        // Serialized strings, not values: key order is part of the contract.
+        assert_eq!(
+            serde_json::to_string(&answers).unwrap(),
+            serde_json::to_string(response.get("answers").unwrap()).unwrap(),
+            "{name}: answers"
+        );
+        assert_eq!(
+            serde_json::to_value(&usage).unwrap(),
+            to_value(response.get("usage").unwrap()),
+            "{name}: usage"
+        );
+        match ran.iter_mut().find(|(c, _)| c == checkpoint) {
+            Some((_, n)) => *n += 1,
+            None => ran.push((checkpoint.to_string(), 1)),
+        }
+    }
+    ran
+}
+
+#[test]
+fn every_english_fixture_matches_the_reference() {
+    let ran = replay(|_| true);
+    let count = |c: &str| ran.iter().find(|(n, _)| n == c).map_or(0, |(_, n)| *n);
+    // The corpus routes 19 cases to English and 1 to typed-decisions; a
+    // drop means a case silently stopped running.
+    assert_eq!(count("english"), 19, "{ran:?}");
+    assert_eq!(count("typed-decisions"), 1, "{ran:?}");
+    if std::env::var_os("LAYA_ML_TOKENIZER_DIR").is_some() {
+        assert_eq!(count("multilingual"), 8, "{ran:?}");
+    }
+}
+
+#[test]
+fn empty_questions_need_no_forward() {
+    let reference = reference();
+    let tok = tokenizer_for("english").unwrap();
+    let cfg = config_for(&reference, "english");
+    let prepared = prepare(
+        &Json::String("x".into()),
+        &[],
+        &cfg,
+        &tok,
+        Budget::default(),
+    )
+    .unwrap();
+    assert!(prepared.rows.is_empty());
+    let (answers, usage) = shape(&prepared, &[], &cfg, None);
+    assert_eq!(serde_json::to_string(&answers).unwrap(), "{}");
+    assert_eq!(
+        serde_json::to_string(&usage).unwrap(),
+        r#"{"input_tokens":0,"output_tokens":0}"#
+    );
+}
+
+#[test]
+fn options_that_cannot_all_fit_are_refused() {
+    let reference = reference();
+    let tok = tokenizer_for("english").unwrap();
+    let cfg = config_for(&reference, "english");
+    let questions = Json::parse(
+        br#"{"q": {"type": "choice", "instructions": "i", "criteria": ["a", "b", "c", "d"]}}"#,
+    )
+    .unwrap();
+    let Json::Object(questions) = questions else {
+        unreachable!()
+    };
+    let budget = Budget {
+        max_len: Some(10),
+        head_max_len: None,
+    };
+    let err = prepare(&Json::String("x".into()), &questions, &cfg, &tok, budget).unwrap_err();
+    assert_eq!(
+        (err.status, err.detail.as_str()),
+        (422, "question 'q' options exceed head_max_len=192")
+    );
+}

crates/neuron/src/harness/decision/question.rs

@@ -0,0 +1,458 @@
+//! Question validation, normalisation and option rendering.
+//!
+//! Every rejection here is a 422 whose message names the question and what
+//! to fix — the same text the reference returns, so a client written
+//! against it sees the same errors.
+
+use super::{DecisionRequestError, unprocessable};
+use cortex_core::decisions::{HashKey, Json};
+use std::collections::HashMap;
+
+/// The three decision primitives. The discriminant is the index the model's
+/// type embedding and the calibration tables use.
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub enum QType {
+    /// Pick one of the caller's labels.
+    Choice = 0,
+    /// An ordinal level on the caller's rubric.
+    Score = 1,
+    /// Yes/no: the probability that a statement holds.
+    Noul = 2,
+}
+
+impl QType {
+    pub const NAMES: [&'static str; 3] = ["choice", "score", "noul"];
+
+    pub fn name(self) -> &'static str {
+        Self::NAMES[self as usize]
+    }
+
+    fn parse(s: &str) -> Option<Self> {
+        match s {
+            "choice" => Some(Self::Choice),
+            "score" => Some(Self::Score),
+            "noul" => Some(Self::Noul),
+            _ => None,
+        }
+    }
+}
+
+/// A validated question, normalised the way the reference normalises it.
+#[derive(Debug, Clone, PartialEq)]
+pub struct Question {
+    pub id: String,
+    pub kind: QType,
+    /// The instruction text: the caller's string, or the JSON rendering of
+    /// a structured value.
+    pub instructions: String,
+    /// Each option's rendered text, in answer-index order. For `noul` this
+    /// is always `[false, true]`, so index 1 is P(true).
+    pub options: Vec<String>,
+    /// `choice`: each option's label as the caller wrote it (the answer key).
+    /// `score`: each level's description (the legend). `noul`: empty.
+    pub labels: Vec<Json>,
+}
+
+impl Question {
+    /// Validate one question definition and normalise it.
+    pub fn from_definition(id: &str, def: &Json) -> Result<Self, DecisionRequestError> {
+        check_question(id, def)?;
+        Ok(to_internal(id, def))
+    }
+}
+
+/// Validate and normalise every question, in request order. The first
+/// invalid question is reported; nothing is normalised until all pass.
+pub fn questions_from_request(
+    questions: &[(String, Json)],
+) -> Result<Vec<Question>, DecisionRequestError> {
+    for (id, def) in questions {
+        check_question(id, def)?;
+    }
+    Ok(questions
+        .iter()
+        .map(|(id, def)| to_internal(id, def))
+        .collect())
+}
+
+fn q(id: &str) -> String {
+    Json::String(id.to_string()).py_repr()
+}
+
+fn check_question(id: &str, def: &Json) -> Result<(), DecisionRequestError> {
+    let qid = q(id);
+    if !matches!(def, Json::Object(_)) {
+        return Err(unprocessable(format!(
+            "question {qid}: definition must be a dict, got {}",
+            def.py_type_name()
+        )));
+    }
+    let t = def.get("type").unwrap_or(&Json::Null);
+    let Some(kind) = t.as_str().and_then(QType::parse) else {
+        return Err(unprocessable(format!(
+            "question {qid}: unknown type {}; use one of ['choice', 'noul', 'score']",
+            t.py_repr()
+        )));
+    };
+    if def.get("instructions").is_none() {
+        return Err(unprocessable(format!(
+            "question {qid}: no 'instructions'; add the text the model should answer"
+        )));
+    }
+    let crit = def.get("criteria");
+    match kind {
+        QType::Choice => check_choice(&qid, crit)?,
+        QType::Score => {
+            let Some(Json::Array(levels)) = crit else {
+                return Err(unprocessable(format!(
+                    "question {qid}: a score question takes 'criteria' as a list of level \
+                     descriptions, index 0 first"
+                )));
+            };
+            if levels.is_empty() {
+                return Err(unprocessable(format!(
+                    "question {qid}: a score question needs at least one level"
+                )));
+            }
+            if let Some(i) = levels.iter().position(|l| *l == Json::Null) {
+                return Err(unprocessable(format!(
+                    "question {qid}: score level {i} is null; give every level a description, \
+                     index 0 first"
+                )));
+            }
+        }
+        QType::Noul => match crit {
+            None | Some(Json::Null) => {}
+            Some(Json::Object(m)) => {
+                let mut keys: Vec<String> = m.iter().map(|(k, _)| k.to_lowercase()).collect();
+                keys.sort();
+                keys.dedup();
+                if keys.iter().any(|k| k != "true" && k != "false") {
+                    let listed = Json::Array(keys.into_iter().map(Json::String).collect());
+                    return Err(unprocessable(format!(
+                        "question {qid}: a noul question takes 'criteria' keyed only \
+                         'true'/'false' (either or both, and omitted is fine), got {}. Those \
+                         keys are the option texts the model reads; any other key was silently \
+                         dropped and replaced with the defaults. If you want the answer worded \
+                         differently, keep 'criteria' keyed 'true'/'false' and set 'labels' \
+                         instead.",
+                        listed.py_repr()
+                    )));
+                }
+            }
+            Some(_) => {
+                return Err(unprocessable(format!(
+                    "question {qid}: a noul question takes 'criteria' as a dict with optional \
+                     'true'/'false' descriptions, or omits it"
+                )));
+            }
+        },
+    }
+    if let Some(labels) = def.get("labels") {
+        if kind != QType::Noul {
+            return Err(unprocessable(format!(
+                "question {qid}: 'labels' is only supported for noul questions"
+            )));
+        }
+        resolve_noul_labels(labels).map_err(|e| unprocessable(format!("question {qid}: {e}")))?;
+    }
+    Ok(())
+}
+
+fn check_choice(qid: &str, crit: Option<&Json>) -> Result<(), DecisionRequestError> {
+    let labels: Vec<Json> = match crit {
+        Some(Json::Object(m)) => m.iter().map(|(k, _)| Json::String(k.clone())).collect(),
+        Some(Json::Array(a)) => a.clone(),
+        _ => {
+            return Err(unprocessable(format!(
+                "question {qid}: a choice question takes 'criteria' as a dict of label -> \
+                 description, or a list of labels"
+            )));
+        }
+    };
+    if labels.is_empty() {
+        return Err(unprocessable(format!(
+            "question {qid}: a choice question needs at least one criterion"
+        )));
+    }
+    for (i, label) in labels.iter().enumerate() {
+        match label {
+            Json::Array(_) | Json::Object(_) => {
+                return Err(unprocessable(format!(
+                    "question {qid}: choice label {i} is a {}; a label is rendered as option \
+                     text and used as the answer key, so it must be a scalar (a string, number \
+                     or bool), got {}",
+                    label.py_type_name(),
+                    label.py_repr()
+                )));
+            }
+            Json::Null => {
+                return Err(unprocessable(format!(
+                    "question {qid}: choice label {i} is null; a label is rendered as option \
+                     text and used as the answer key, so it must be a string, number or bool \
+                     -- a null label renders as the text \"None\" while its answer key is \
+                     \"null\""
+                )));
+            }
+            _ => {}
+        }
+    }
+    // Only the list form can repeat a key: an object's keys are unique.
+    if matches!(crit, Some(Json::Array(_))) {
+        let mut seen: HashMap<HashKey, usize> = HashMap::new();
+        for (i, label) in labels.iter().enumerate() {
+            let key = label.hash_key().expect("scalars are hashable");
+            if key == HashKey::Unique {
+                continue;
+            }
+            if let Some(first) = seen.get(&key) {
+                return Err(unprocessable(format!(
+                    "question {qid}: choice label {i} ({}) repeats label {first}; the labels are \
+                     the answer keys, so every option needs its own (1, 1.0 and True are one key)",
+                    label.py_repr()
+                )));
+            }
+            seen.insert(key, i);
+        }
+    }
+    Ok(())
+}
+
+const DEFAULT_NOUL_FALSE: &str = "no, the statement does not hold";
+const DEFAULT_NOUL_TRUE: &str = "yes, the statement holds";
+
+/// The words a noul question shows the model for its two outcomes.
+fn resolve_noul_labels(labels: &Json) -> Result<(String, String), &'static str> {
+    const BAD: &str =
+        "noul labels must map exactly 'false' and 'true' to distinct non-empty strings";
+    let Json::Object(m) = labels else {
+        return if *labels == Json::Null {
+            Ok(("false".into(), "true".into()))
+        } else {
+            Err(BAD)
+        };
+    };
+    let mut keys: Vec<&str> = m.iter().map(|(k, _)| k.as_str()).collect();
+    keys.sort_unstable();
+    if keys != ["false", "true"] {
+        return Err(BAD);
+    }
+    let (Some(Json::String(f)), Some(Json::String(t))) = (labels.get("false"), labels.get("true"))
+    else {
+        return Err(BAD);
+    };
+    let (f, t) = (py_strip(f), py_strip(t));
+    if f.is_empty() || t.is_empty() || f == t {
+        return Err(BAD);
+    }
+    Ok((f.to_string(), t.to_string()))
+}
+
+/// Python's `str.strip()` with no argument: Unicode whitespace.
+fn py_strip(s: &str) -> &str {
+    s.trim_matches(|c: char| c.is_whitespace() || ('\u{1c}'..='\u{1f}').contains(&c))
+}
+
+/// `render_criterion`: a string as itself, anything else as JSON.
+fn render_criterion(v: &Json) -> String {
+    match v {
+        Json::String(s) => s.clone(),
+        other => other.dumps(),
+    }
+}
+
+/// Only `None` and `""` mean "no description"; `0` and `false` are values.
+fn is_blank(v: &Json) -> bool {
+    matches!(v, Json::Null) || matches!(v, Json::String(s) if s.is_empty())
+}
+
+fn to_internal(id: &str, def: &Json) -> Question {
+    let kind = def
+        .get("type")
+        .and_then(Json::as_str)
+        .and_then(QType::parse)
+        .expect("validated");
+    let instructions = match def.get("instructions").expect("validated") {
+        Json::String(s) => s.clone(),
+        other => other.dumps(),
+    };
+    let crit = def.get("criteria");
+    let (options, labels) = match kind {
+        QType::Choice => {
+            let pairs: Vec<(Json, Json)> = match crit {
+                Some(Json::Object(m)) => m
+                    .iter()
+                    .map(|(k, v)| (Json::String(k.clone()), v.clone()))
+                    .collect(),
+                Some(Json::Array(a)) => a.iter().map(|l| (l.clone(), Json::Null)).collect(),
+                _ => unreachable!("validated"),
+            };
+            let options = pairs
+                .iter()
+                .map(|(k, v)| {
+                    if is_blank(v) {
+                        k.py_str()
+                    } else {
+                        format!("{}: {}", k.py_str(), render_criterion(v))
+                    }
+                })
+                .collect();
+            (options, pairs.into_iter().map(|(k, _)| k).collect())
+        }
+        QType::Score => {
+            let Some(Json::Array(levels)) = crit else {
+                unreachable!("validated")
+            };
+            let options = levels
+                .iter()
+                .enumerate()
+                .map(|(i, c)| format!("level {i}: {}", render_criterion(c)))
+                .collect();
+            (options, levels.clone())
+        }
+        QType::Noul => {
+            // Keys are matched case-insensitively; a repeat after
+            // lower-casing keeps the later value, as a dict would.
+            let mut desc: HashMap<String, Json> = HashMap::new();
+            if let Some(Json::Object(m)) = crit {
+                for (k, v) in m {
+                    desc.insert(k.to_lowercase(), v.clone());
+                }
+            }
+            let (fl, tl) =
+                resolve_noul_labels(def.get("labels").unwrap_or(&Json::Null)).expect("validated");
+            let render = |key: &str, default: &str| match desc.get(key) {
+                Some(v) if !is_blank(v) => render_criterion(v),
+                _ => default.to_string(),
+            };
+            (
+                vec![
+                    format!("{fl}: {}", render("false", DEFAULT_NOUL_FALSE)),
+                    format!("{tl}: {}", render("true", DEFAULT_NOUL_TRUE)),
+                ],
+                Vec::new(),
+            )
+        }
+    };
+    Question {
+        id: id.to_string(),
+        kind,
+        instructions,
+        options,
+        labels,
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn def(s: &str) -> Json {
+        Json::parse(s.as_bytes()).unwrap()
+    }
+
+    fn err(s: &str) -> String {
+        Question::from_definition("q", &def(s)).unwrap_err().detail
+    }
+
+    #[test]
+    fn options_render_as_the_reference_renders_them() {
+        let c = Question::from_definition(
+            "q",
+            &def(r#"{"type": "choice", "instructions": {"a": 1},
+                     "criteria": {"x": "", "y": null, "z": 0, "w": {"d": [1, 2.0]}, "v": "text"}}"#),
+        )
+        .unwrap();
+        assert_eq!(c.instructions, r#"{"a": 1}"#);
+        assert_eq!(
+            c.options,
+            ["x", "y", "z: 0", r#"w: {"d": [1, 2.0]}"#, "v: text"]
+        );
+
+        let l = Question::from_definition(
+            "q",
+            &def(r#"{"type": "choice", "instructions": "i", "criteria": [1, 2.5, false, "s"]}"#),
+        )
+        .unwrap();
+        assert_eq!(l.options, ["1", "2.5", "False", "s"]);
+
+        let s = Question::from_definition(
+            "q",
+            &def(r#"{"type": "score", "instructions": "i", "criteria": ["calm", {"k": false}]}"#),
+        )
+        .unwrap();
+        assert_eq!(s.options, ["level 0: calm", r#"level 1: {"k": false}"#]);
+
+        let n = Question::from_definition(
+            "q",
+            &def(
+                r#"{"type": "noul", "instructions": "i", "criteria": {"TRUE": "held", "false": ""},
+                     "labels": {"false": " no ", "true": "yes"}}"#,
+            ),
+        )
+        .unwrap();
+        assert_eq!(
+            n.options,
+            ["no: no, the statement does not hold", "yes: held"]
+        );
+    }
+
+    #[test]
+    fn invalid_questions_name_the_question_and_the_fix() {
+        assert_eq!(
+            err("[]"),
+            "question 'q': definition must be a dict, got list"
+        );
+        assert_eq!(
+            err(r#"{"type": "pick"}"#),
+            "question 'q': unknown type 'pick'; use one of ['choice', 'noul', 'score']"
+        );
+        assert_eq!(
+            err(r#"{"type": "noul"}"#),
+            "question 'q': no 'instructions'; add the text the model should answer"
+        );
+        assert!(
+            err(r#"{"type": "choice", "instructions": "i", "criteria": []}"#)
+                .ends_with("needs at least one criterion")
+        );
+        assert!(
+            err(r#"{"type": "choice", "instructions": "i", "criteria": [[1]]}"#)
+                .starts_with("question 'q': choice label 0 is a list;")
+        );
+        assert!(
+            err(r#"{"type": "choice", "instructions": "i", "criteria": ["a", null]}"#)
+                .starts_with("question 'q': choice label 1 is null;")
+        );
+        assert!(
+            err(r#"{"type": "choice", "instructions": "i", "criteria": [1, "x", true]}"#)
+                .starts_with("question 'q': choice label 2 (True) repeats label 0;")
+        );
+        assert_eq!(
+            err(r#"{"type": "score", "instructions": "i", "criteria": ["a", null]}"#),
+            "question 'q': score level 1 is null; give every level a description, index 0 first"
+        );
+        assert!(
+            err(r#"{"type": "noul", "instructions": "i", "criteria": {"yes": "x"}}"#)
+                .contains("got ['yes'].")
+        );
+        assert_eq!(
+            err(r#"{"type": "score", "instructions": "i", "criteria": ["a"], "labels": {}}"#),
+            "question 'q': 'labels' is only supported for noul questions"
+        );
+        assert_eq!(
+            err(r#"{"type": "noul", "instructions": "i", "labels": {"false": "x", "true": " x"}}"#),
+            "question 'q': noul labels must map exactly 'false' and 'true' to distinct non-empty strings"
+        );
+    }
+
+    #[test]
+    fn nan_labels_never_collide() {
+        assert!(
+            Question::from_definition(
+                "q",
+                &def(r#"{"type": "choice", "instructions": "i", "criteria": [NaN, NaN]}"#),
+            )
+            .is_ok()
+        );
+    }
+}

crates/neuron/src/harness/decision/route.rs

@@ -0,0 +1,301 @@
+//! Choosing which checkpoint of a Laya family answers a request (#339).
+//!
+//! A released Laya repo carries three checkpoints: the English one at
+//! its root, `multilingual/` and `typed-decisions/`. The English
+//! checkpoint cannot read other scripts — upstream measured Khmer at
+//! 0.000 accuracy with 0.952 confidence — so an unpinned request is
+//! routed by the language of its state. This follows the reference's
+//! `Router._route` (`laya/router.py`) and its HTTP server's
+//! `_resolve_model` (`laya/serve.py`), including the `reason` strings.
+//!
+//! Precedence, as upstream: a checkpoint the caller names in `model`,
+//! then the detected script and language, then the default (English).
+//! Upstream's opt-in typed-decisions workflow routing and caller
+//! language hints are not exposed over HTTP by the reference, so they
+//! are not here either; the matched workflow is still reported.
+
+use cortex_core::decisions::{DecisionRouting, Json};
+
+use crate::harness::lang_detect;
+
+/// The root checkpoint's routing name.
+pub const ENGLISH: &str = "english";
+pub const MULTILINGUAL: &str = "multilingual";
+pub const TYPED_DECISIONS: &str = "typed-decisions";
+
+/// Where an unpinned request with no usable language signal goes.
+pub const DEFAULT_CHECKPOINT: &str = ENGLISH;
+
+/// Every checkpoint a released family carries, root first.
+pub const FAMILY: [&str; 3] = [ENGLISH, MULTILINGUAL, TYPED_DECISIONS];
+
+/// The reference's `_ALIASES`: names people are likely to type.
+const ALIASES: [(&str, &str); 10] = [
+    ("en", ENGLISH),
+    ("laya", ENGLISH),
+    ("default", ENGLISH),
+    ("multi", MULTILINGUAL),
+    ("ml", MULTILINGUAL),
+    ("laya-multilingual", MULTILINGUAL),
+    ("typed", TYPED_DECISIONS),
+    ("typed_decisions", TYPED_DECISIONS),
+    ("laya-typed-decisions", TYPED_DECISIONS),
+    ("decisions", TYPED_DECISIONS),
+];
+
+/// The reference server's `_PUBLISHED_MODEL_IDS`: the Hugging Face ids
+/// the non-root checkpoints are published under.
+const PUBLISHED: [(&str, &str); 2] = [
+    ("convaiinnovations/laya-multilingual", MULTILINGUAL),
+    ("convaiinnovations/laya-typed-decisions", TYPED_DECISIONS),
+];
+
+/// Question-id signatures of the four typed-decisions workflows.
+const WORKFLOWS: [(&str, [&str; 5]); 4] = [
+    (
+        "agent_trace_observability",
+        ["action", "needs_review", "outcome", "risk", "urgency"],
+    ),
+    (
+        "customer_service",
+        ["action", "category", "churn_risk", "needs_human", "urgency"],
+    ),
+    (
+        "invoice_processing",
+        [
+            "discrepancy_severity",
+            "disposition",
+            "duplicate",
+            "matches_order",
+            "urgency",
+        ],
+    ),
+    (
+        "security_incidents",
+        [
+            "credential_compromise",
+            "disposition",
+            "severity",
+            "true_positive",
+            "urgency",
+        ],
+    ),
+];
+
+/// The checkpoint a request's `model` field pins, if it names one.
+///
+/// Anything else — a Jev service name like `jev-1`, the family's own
+/// repo id, an empty string — pins nothing and the request is routed.
+pub fn resolve_pin(model: Option<&str>) -> Option<&'static str> {
+    let model = model?;
+    if model.is_empty() {
+        return None;
+    }
+    let key = py_strip(model).to_lowercase();
+    if let Some((_, ckpt)) = PUBLISHED.iter().find(|(id, _)| *id == key) {
+        return Some(ckpt);
+    }
+    let key = ALIASES
+        .iter()
+        .find(|(alias, _)| *alias == key)
+        .map(|(_, ckpt)| *ckpt)
+        .unwrap_or(key.as_str());
+    FAMILY.iter().copied().find(|c| *c == key)
+}
+
+/// The typed-decisions workflow whose question ids these are exactly.
+pub fn match_workflow(questions: &[(String, Json)]) -> Option<&'static str> {
+    let ids: std::collections::BTreeSet<&str> = questions.iter().map(|(q, _)| q.as_str()).collect();
+    WORKFLOWS.iter().find_map(|(name, sig)| {
+        let sig: std::collections::BTreeSet<&str> = sig.iter().copied().collect();
+        (ids == sig).then_some(*name)
+    })
+}
+
+/// Decide which checkpoint answers, and say why.
+///
+/// `model_id` is the served family's repo id, used for the `repo` field.
+/// The result names a checkpoint whether or not this node has it
+/// loaded; the caller decides what to do about one it lacks.
+pub fn route(
+    model_id: &str,
+    state: &Json,
+    questions: &[(String, Json)],
+    pin: Option<&'static str>,
+) -> DecisionRouting {
+    if let Some(key) = pin {
+        return DecisionRouting {
+            model: key.to_string(),
+            repo: repo_of(model_id, key),
+            reason: format!("explicit model={}", py_repr(key)),
+            detection: None,
+            workflow: None,
+        };
+    }
+    let workflow = match_workflow(questions).map(str::to_string);
+    let det = lang_detect::analyse(state);
+    let (key, reason) = if det.script == "unknown" {
+        (
+            DEFAULT_CHECKPOINT,
+            format!("no letters detected in state; using default ({DEFAULT_CHECKPOINT})"),
+        )
+    } else if det.script != "latin" {
+        (
+            MULTILINGUAL,
+            format!(
+                "non-Latin script ({}, {:.0}% of letters); the English checkpoint cannot read it",
+                det.script,
+                100.0 * det.non_latin_fraction
+            ),
+        )
+    } else if !det.is_english {
+        let reason = if let Some(segment) = &det.mixed_segment {
+            format!(
+                "Latin script, mostly English, but a line or field reads as {} ({}); \
+                 the English checkpoint cannot read it",
+                det.language.map(py_repr).unwrap_or_else(|| "None".into()),
+                py_repr(&segment.chars().take(60).collect::<String>()),
+            )
+        } else if let Some(lang) = det.language {
+            format!(
+                "Latin script but language looks like {}, not English",
+                py_repr(lang)
+            )
+        } else {
+            format!(
+                "Latin script, language not identified but {:.0}% non-English letters; \
+                 not safe for the English checkpoint",
+                100.0 * det.diacritic_rate
+            )
+        };
+        (MULTILINGUAL, reason)
+    } else if det.language_undecided {
+        (
+            DEFAULT_CHECKPOINT,
+            format!(
+                "Latin script, language not identified and no non-English letters; \
+                 using default ({DEFAULT_CHECKPOINT})"
+            ),
+        )
+    } else {
+        (ENGLISH, "English Latin text".to_string())
+    };
+    DecisionRouting {
+        model: key.to_string(),
+        repo: repo_of(model_id, key),
+        reason,
+        detection: Some(analysis_json(&det)),
+        workflow,
+    }
+}
+
+/// The language per-language calibration keys on: the detected one.
+pub fn detected_language(routing: &DecisionRouting) -> Option<&str> {
+    routing
+        .detection
+        .as_ref()
+        .and_then(|d| d.get("language"))
+        .and_then(Json::as_str)
+}
+
+/// `<repo>` for the root checkpoint, `<repo>/<subfolder>` otherwise.
+pub fn repo_of(model_id: &str, checkpoint: &str) -> String {
+    if checkpoint == ENGLISH {
+        model_id.to_string()
+    } else {
+        format!("{model_id}/{checkpoint}")
+    }
+}
+
+/// The analysis as an order-preserving [`Json`], so the response keeps
+/// the reference's field order.
+fn analysis_json(det: &lang_detect::Analysis) -> Json {
+    let text = serde_json::to_string(det).unwrap_or_else(|_| "null".into());
+    Json::parse(text.as_bytes()).unwrap_or(Json::Null)
+}
+
+fn py_repr(s: &str) -> String {
+    Json::String(s.to_string()).py_repr()
+}
+
+/// Python's `str.strip()`: Unicode whitespace, which includes the
+/// U+001C–U+001F separators Rust's `trim` leaves alone.
+fn py_strip(s: &str) -> &str {
+    s.trim_matches(|c: char| c.is_whitespace() || ('\u{1c}'..='\u{1f}').contains(&c))
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn pins_follow_the_reference_server() {
+        assert_eq!(resolve_pin(Some("multilingual")), Some(MULTILINGUAL));
+        assert_eq!(resolve_pin(Some(" ML ")), Some(MULTILINGUAL));
+        assert_eq!(
+            resolve_pin(Some("convaiinnovations/laya-typed-decisions")),
+            Some(TYPED_DECISIONS)
+        );
+        assert_eq!(resolve_pin(Some("laya")), Some(ENGLISH));
+        // A Jev service name, the family's own id, and nothing: no pin.
+        assert_eq!(resolve_pin(Some("jev-1")), None);
+        assert_eq!(resolve_pin(Some("convaiinnovations/laya")), None);
+        assert_eq!(resolve_pin(Some("")), None);
+        assert_eq!(resolve_pin(None), None);
+    }
+
+    #[test]
+    fn workflow_needs_the_exact_id_set() {
+        let qs = |ids: &[&str]| -> Vec<(String, Json)> {
+            ids.iter().map(|i| (i.to_string(), Json::Null)).collect()
+        };
+        assert_eq!(
+            match_workflow(&qs(&[
+                "urgency",
+                "action",
+                "category",
+                "churn_risk",
+                "needs_human"
+            ])),
+            Some("customer_service")
+        );
+        assert_eq!(match_workflow(&qs(&["urgency"])), None);
+    }
+
+    /// Every case in the reference fixtures routes to the checkpoint, with
+    /// the reason and detection, the reference reported.
+    #[test]
+    fn routes_every_reference_case_like_the_reference() {
+        let text = std::fs::read_to_string(
+            std::path::Path::new(env!("CARGO_MANIFEST_DIR"))
+                .join("src/harness/testdata/laya/reference.json"),
+        )
+        .expect("read reference.json");
+        let reference = Json::parse(text.as_bytes()).expect("parse reference.json");
+        let Some(Json::Array(cases)) = reference.get("cases") else {
+            panic!("no cases");
+        };
+        for case in cases {
+            let name = case.get("name").and_then(Json::as_str).unwrap();
+            let request = case.get("request").unwrap();
+            let body = serde_json::to_vec(request).unwrap();
+            let req = cortex_core::decisions::SystemOneRequest::parse(
+                &body,
+                cortex_core::decisions::limits::DEFAULT_MAX_TOKEN_BUDGET,
+            )
+            .unwrap();
+            let routing = route(
+                "convaiinnovations/laya",
+                &req.state,
+                &req.questions,
+                resolve_pin(req.model.as_deref()),
+            );
+            let want = case.get("response").unwrap().get("routing").unwrap();
+            assert_eq!(
+                serde_json::to_string(&routing).unwrap(),
+                serde_json::to_string(want).unwrap(),
+                "{name}"
+            );
+        }
+    }
+}

crates/neuron/src/harness/decision/sequence.rs

@@ -0,0 +1,316 @@
+//! Sequence assembly: one token sequence per question, under the model's
+//! token budgets.
+//!
+//! ```text
+//! [CLS] "<type> question: <instructions>" [SEP] [MASK] opt0 [MASK] opt1 … [SEP] state [SEP]
+//! ```
+//!
+//! The budget rules, in the order they apply:
+//!
+//! 1. Each option is `[MASK]` plus at most 48 tokens of its text.
+//! 2. `head_max_len` bounds the instruction text and all options together.
+//!    If fewer than 16 tokens would remain for the instructions, every
+//!    option is cut to the same `max(4, (head_max_len - 16) / options)`
+//!    tokens (marker included).
+//! 3. The instructions get what is left of `head_max_len`, but never fewer
+//!    than 8 tokens.
+//! 4. The state fills the rest of `max_len`. A conversation (a JSON list)
+//!    keeps its newest turns, so it is cut from the left; anything else is
+//!    cut from the right.
+//! 5. The whole sequence is cut to `max_len`, dropping any marker that no
+//!    longer fits — a question whose markers do not all fit is refused.
+
+use super::question::{QType, Question, questions_from_request};
+use super::tokenizer::TextEncoder;
+use super::{DecisionConfig, DecisionRequestError, unprocessable};
+use cortex_core::decisions::Json;
+
+/// Option text tokens kept before the head budget is considered.
+pub const MAX_OPTION_TOKENS: usize = 48;
+/// Head tokens the instructions are guaranteed before options are shrunk.
+const MIN_INSTRUCTION_ROOM: usize = 16;
+/// Per-option floor (marker included) when options are shrunk.
+const MIN_TOKENS_PER_OPTION: usize = 4;
+/// Instruction tokens kept however many the options take.
+const MIN_INSTRUCTION_TOKENS: usize = 8;
+
+/// Per-request overrides of the checkpoint's token budgets.
+#[derive(Debug, Clone, Copy, Default, PartialEq, Eq)]
+pub struct Budget {
+    pub max_len: Option<usize>,
+    pub head_max_len: Option<usize>,
+}
+
+/// One model input row: one question over the state.
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct Row {
+    pub input_ids: Vec<u32>,
+    /// Position of each option's `[MASK]` marker, in option order.
+    pub markers: Vec<usize>,
+    pub qtype: QType,
+}
+
+/// What the head budget did to a question's options.
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub struct OptionStats {
+    pub options: usize,
+    /// Options whose token span is still unique after the cut. Fewer than
+    /// `options` means the model can no longer tell some of them apart.
+    pub options_distinct: usize,
+    /// The per-option cap applied, when the options had to be shrunk.
+    pub tokens_per_option: Option<usize>,
+}
+
+/// A request turned into model input: one [`Row`] per question, in request
+/// order. With no questions there are no rows and nothing to run.
+#[derive(Debug, Clone, PartialEq)]
+pub struct Prepared {
+    pub questions: Vec<Question>,
+    pub rows: Vec<Row>,
+    pub stats: Vec<OptionStats>,
+    /// Every token the forward pass encodes: the metered input.
+    pub input_tokens: usize,
+    /// The id batched rows are padded with.
+    pub pad_id: u32,
+}
+
+/// The text a state is tokenized from: a string as-is, anything else as
+/// `json.dumps(state, ensure_ascii=False)`.
+pub fn serialize_state(state: &Json) -> String {
+    match state {
+        Json::String(s) => s.clone(),
+        other => other.dumps(),
+    }
+}
+
+/// Validate a request's questions and build their model input.
+///
+/// `questions` are the request's definitions in request order; `cfg` and
+/// `tok` belong to the checkpoint the request was routed to.
+pub fn prepare(
+    state: &Json,
+    questions: &[(String, Json)],
+    cfg: &DecisionConfig,
+    tok: &dyn TextEncoder,
+    budget: Budget,
+) -> Result<Prepared, DecisionRequestError> {
+    let special = tok.special();
+    if questions.is_empty() {
+        return Ok(Prepared {
+            questions: Vec::new(),
+            rows: Vec::new(),
+            stats: Vec::new(),
+            input_tokens: 0,
+            pad_id: special.pad_id,
+        });
+    }
+    let questions = questions_from_request(questions)?;
+    let max_len = budget.max_len.unwrap_or(cfg.max_len);
+    let head_max_len = budget.head_max_len.unwrap_or(cfg.head_max_len);
+    let encode = |text: &str| {
+        tok.encode(text)
+            .map_err(|e| DecisionRequestError::new(500, format!("tokenizer failed: {e}")))
+    };
+    // Tokenized once and sliced per question.
+    let state_ids = encode(&serialize_state(state).replace(&special.mask_token, " "))?;
+    let truncate_left = matches!(state, Json::Array(_));
+    let frame = Frame {
+        cls: special.cls_id,
+        sep: special.sep_id,
+        max_len,
+        head_max_len,
+    };
+
+    let mut rows = Vec::with_capacity(questions.len());
+    let mut stats = Vec::with_capacity(questions.len());
+    for question in &questions {
+        let instructions = question.instructions.replace(&special.mask_token, " ");
+        let head_ids = encode(&format!(
+            "{} question: {instructions}",
+            question.kind.name()
+        ))?;
+        let mut option_ids = Vec::with_capacity(question.options.len());
+        for option in &question.options {
+            let mut ids = vec![special.mask_id];
+            let text = encode(&format!(" {}", option.replace(&special.mask_token, " ")))?;
+            ids.extend(text.into_iter().take(MAX_OPTION_TOKENS));
+            option_ids.push(ids);
+        }
+        let (row, s) = frame.assemble(head_ids, option_ids, &state_ids, truncate_left);
+        if row.markers.len() != question.options.len() {
+            return Err(unprocessable(format!(
+                "question {} options exceed head_max_len={head_max_len}",
+                Json::String(question.id.clone()).py_repr()
+            )));
+        }
+        rows.push(Row {
+            qtype: question.kind,
+            ..row
+        });
+        stats.push(s);
+    }
+    let input_tokens = rows.iter().map(|r| r.input_ids.len()).sum();
+    Ok(Prepared {
+        questions,
+        rows,
+        stats,
+        input_tokens,
+        pad_id: special.pad_id,
+    })
+}
+
+/// The fixed parts of every sequence for one request.
+struct Frame {
+    cls: u32,
+    sep: u32,
+    max_len: usize,
+    head_max_len: usize,
+}
+
+impl Frame {
+    /// The pure budget arithmetic of one sequence, over already-tokenized parts.
+    fn assemble(
+        &self,
+        mut head_ids: Vec<u32>,
+        mut option_ids: Vec<Vec<u32>>,
+        state_ids: &[u32],
+        truncate_left: bool,
+    ) -> (Row, OptionStats) {
+        let Frame {
+            cls,
+            sep,
+            max_len,
+            head_max_len,
+        } = *self;
+        let used = |o: &[Vec<u32>]| o.iter().map(Vec::len).sum::<usize>();
+        // Signed: options can overrun the head budget.
+        let mut opt_budget = head_max_len as isize - used(&option_ids) as isize;
+        let mut tokens_per_option = None;
+        if opt_budget < MIN_INSTRUCTION_ROOM as isize {
+            let per = MIN_TOKENS_PER_OPTION
+                .max(head_max_len.saturating_sub(MIN_INSTRUCTION_ROOM) / option_ids.len().max(1));
+            tokens_per_option = Some(per);
+            for o in &mut option_ids {
+                o.truncate(per);
+            }
+            opt_budget = head_max_len as isize - used(&option_ids) as isize;
+        }
+        head_ids.truncate(opt_budget.max(MIN_INSTRUCTION_TOKENS as isize) as usize);
+
+        let mut ids = Vec::with_capacity(max_len);
+        ids.push(cls);
+        ids.extend(head_ids);
+        ids.push(sep);
+        let mut markers = Vec::with_capacity(option_ids.len());
+        for o in &option_ids {
+            markers.push(ids.len());
+            ids.extend_from_slice(o);
+        }
+        ids.push(sep);
+        let room = max_len.saturating_sub(ids.len() + 1);
+        let state = if truncate_left {
+            &state_ids[state_ids.len().saturating_sub(room)..]
+        } else {
+            &state_ids[..room.min(state_ids.len())]
+        };
+        ids.extend_from_slice(state);
+        ids.push(sep);
+        ids.truncate(max_len);
+        markers.retain(|&m| m < max_len);
+
+        let mut distinct = option_ids.clone();
+        distinct.sort();
+        distinct.dedup();
+        (
+            Row {
+                input_ids: ids,
+                markers,
+                qtype: QType::Choice,
+            },
+            OptionStats {
+                options: option_ids.len(),
+                options_distinct: distinct.len(),
+                tokens_per_option,
+            },
+        )
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    const CLS: u32 = 1;
+    const SEP: u32 = 2;
+    const M: u32 = 9;
+
+    fn frame(max_len: usize, head_max_len: usize) -> Frame {
+        Frame {
+            cls: CLS,
+            sep: SEP,
+            max_len,
+            head_max_len,
+        }
+    }
+
+    fn opts(lens: &[usize]) -> Vec<Vec<u32>> {
+        lens.iter()
+            .enumerate()
+            .map(|(i, &n)| {
+                let mut o = vec![M];
+                o.extend((0..n - 1).map(|j| 100 + (i * 50 + j) as u32));
+                o
+            })
+            .collect()
+    }
+
+    #[test]
+    fn roomy_budget_keeps_everything() {
+        let (row, s) = frame(64, 32).assemble(vec![10, 11], opts(&[2, 3]), &[50, 51], false);
+        assert_eq!(
+            row.input_ids,
+            [CLS, 10, 11, SEP, M, 100, M, 150, 151, SEP, 50, 51, SEP]
+        );
+        assert_eq!(row.markers, [4, 6]);
+        assert_eq!(
+            (s.options, s.options_distinct, s.tokens_per_option),
+            (2, 2, None)
+        );
+    }
+
+    #[test]
+    fn crowded_head_shrinks_options_evenly_and_keeps_eight_instruction_tokens() {
+        // 3 options of 10 = 30 tokens against a 40-token head: 10 left < 16.
+        let head: Vec<u32> = (10..40).collect();
+        let (row, s) = frame(128, 40).assemble(head, opts(&[10, 10, 10]), &[], false);
+        // per = max(4, (40 - 16) / 3) = 8; 40 - 24 = 16 instruction tokens.
+        assert_eq!(s.tokens_per_option, Some(8));
+        assert_eq!(row.markers, [18, 26, 34]);
+        let (row, s) = frame(256, 40).assemble((10..40).collect(), opts(&[10; 12]), &[], false);
+        // per = max(4, 24 / 12 = 2) = 4; the budget goes negative, 8 tokens survive.
+        assert_eq!(s.tokens_per_option, Some(4));
+        assert_eq!(row.markers[0], 1 + 8 + 1);
+    }
+
+    #[test]
+    fn state_is_cut_from_the_side_the_state_shape_asks_for() {
+        let state: Vec<u32> = (50..60).collect();
+        let (right, _) = frame(8, 32).assemble(vec![], opts(&[1]), &state, false);
+        let (left, _) = frame(8, 32).assemble(vec![], opts(&[1]), &state, true);
+        // [CLS] [SEP] [MASK] [SEP] = 4 tokens, 1 reserved for the last [SEP]: room 3.
+        assert_eq!(right.input_ids, [CLS, SEP, M, SEP, 50, 51, 52, SEP]);
+        assert_eq!(left.input_ids, [CLS, SEP, M, SEP, 57, 58, 59, SEP]);
+        // No room at all: none of the state, not all of it.
+        let (none, _) = frame(5, 32).assemble(vec![], opts(&[1]), &state, true);
+        assert_eq!(none.input_ids, [CLS, SEP, M, SEP, SEP]);
+    }
+
+    #[test]
+    fn markers_past_max_len_are_dropped_and_collisions_counted() {
+        let same = vec![vec![M, 7, 7], vec![M, 7, 7]];
+        let (row, s) = frame(5, 64).assemble(vec![], same, &[], false);
+        // [CLS] [SEP] then markers at 2 and 5; only the first fits in 5.
+        assert_eq!(row.markers, [2]);
+        assert_eq!((s.options, s.options_distinct), (2, 1));
+    }
+}

crates/neuron/src/harness/decision/shape.rs

@@ -0,0 +1,318 @@
+//! Per-option logits → calibrated, typed answers.
+//!
+//! The reference computes this in numpy on the float32 logits the model
+//! returns, and publishes each probability rounded to 4 decimal places.
+//! Reproducing the rounded values exactly means reproducing the float32
+//! arithmetic, in numpy's order: division by a float32 temperature, a
+//! max-subtracted `exp`, and numpy's *pairwise* summation (not a running
+//! sum) for every reduction. The expected score is the exception: numpy
+//! promotes `arange(k) * p` to float64.
+
+use super::question::QType;
+use super::sequence::Prepared;
+use super::{DECISION_MODEL_NAME, DecisionConfig};
+use cortex_core::decisions::{
+    CollapsedOptions, DecisionAction, DecisionAnswer, DecisionUsage, OrderedMap, SystemOneResponse,
+};
+
+/// The model's output for one [`super::Row`].
+#[derive(Debug, Clone, PartialEq)]
+pub struct RowOutput {
+    /// The scorer's raw logit at each option marker, before calibration.
+    /// Entries past the row's marker count (batch padding) are ignored.
+    pub logits: Vec<f32>,
+    /// The act head's probability of acting: `softmax(act_logits)[0]`.
+    pub act_probability: f32,
+}
+
+/// Build the answers and usage for a prepared request from the model's
+/// per-row outputs (`outputs[i]` belongs to `prepared.rows[i]`).
+///
+/// `lang` selects per-language calibration where the checkpoint has it:
+/// the language the request was declared or detected as.
+pub fn shape(
+    prepared: &Prepared,
+    outputs: &[RowOutput],
+    cfg: &DecisionConfig,
+    lang: Option<&str>,
+) -> (OrderedMap<DecisionAnswer>, DecisionUsage) {
+    assert_eq!(
+        outputs.len(),
+        prepared.rows.len(),
+        "one model output per prepared row"
+    );
+    let temps = cfg.temperatures_for(lang);
+    let mut answers = Vec::with_capacity(outputs.len());
+    let mut collapsed = Vec::new();
+    for ((question, row), (out, stats)) in prepared
+        .questions
+        .iter()
+        .zip(&prepared.rows)
+        .zip(outputs.iter().zip(&prepared.stats))
+    {
+        let k = row.markers.len();
+        let p = calibrate(
+            &out.logits[..k],
+            temps.for_question(question.kind as usize, k),
+        );
+        let answer_confidence = round4(
+            p.iter()
+                .copied()
+                .fold(f32::NEG_INFINITY, f32::max)
+                .clamp(0.0, 1.0) as f64,
+        );
+        let action = DecisionAction {
+            act_probability: round4(out.act_probability as f64),
+        };
+        let probabilities = |key: &dyn Fn(usize) -> String| {
+            OrderedMap(
+                p.iter()
+                    .enumerate()
+                    .map(|(i, &v)| (key(i), round4(v as f64)))
+                    .collect(),
+            )
+        };
+        let answer = match question.kind {
+            QType::Choice => DecisionAnswer::Choice {
+                choice: question.labels[argmax(&p)].clone(),
+                probabilities: probabilities(&|i| {
+                    question.labels[i]
+                        .json_key()
+                        .expect("choice labels are scalars")
+                }),
+                confidence: round4(entropy_confidence(&p) as f64),
+                answer_confidence,
+                action,
+            },
+            QType::Score => {
+                // int64 * float32 promotes to float64 in numpy.
+                let expected: Vec<f64> = p
+                    .iter()
+                    .enumerate()
+                    .map(|(i, &v)| i as f64 * v as f64)
+                    .collect();
+                DecisionAnswer::Score {
+                    score: round4(pairwise_sum(&expected)),
+                    legend: OrderedMap(
+                        question
+                            .labels
+                            .iter()
+                            .enumerate()
+                            .map(|(i, c)| (i.to_string(), c.clone()))
+                            .collect(),
+                    ),
+                    probabilities: probabilities(&|i| i.to_string()),
+                    confidence: round4(entropy_confidence(&p) as f64),
+                    answer_confidence,
+                    action,
+                }
+            }
+            QType::Noul => {
+                let yes = p[1] as f64;
+                DecisionAnswer::Noul {
+                    noul: round4(yes),
+                    confidence: round4(yes.max(1.0 - yes)),
+                    answer_confidence,
+                    action,
+                }
+            }
+        };
+        answers.push((question.id.clone(), answer));
+        if stats.options_distinct < stats.options {
+            collapsed.push((
+                question.id.clone(),
+                CollapsedOptions {
+                    total: stats.options,
+                    distinct: stats.options_distinct,
+                    tokens_per_option: stats.tokens_per_option,
+                },
+            ));
+        }
+    }
+    let usage = DecisionUsage {
+        input_tokens: prepared.input_tokens,
+        output_tokens: 0,
+        options: (!collapsed.is_empty()).then_some(OrderedMap(collapsed)),
+    };
+    (OrderedMap(answers), usage)
+}
+
+/// Assemble the full response once routing is known.
+pub fn response(
+    answers: OrderedMap<DecisionAnswer>,
+    usage: DecisionUsage,
+    routing: Option<cortex_core::decisions::DecisionRouting>,
+) -> SystemOneResponse {
+    SystemOneResponse {
+        model: DECISION_MODEL_NAME.to_string(),
+        answers,
+        usage,
+        routing,
+    }
+}
+
+/// Temperature-scaled softmax, as the reference evaluates it in numpy:
+/// `z = logits / t`, `e = exp(z - z.max())`, `p = e / e.sum()`.
+///
+/// numpy's SIMD `expf` and libm's disagree by one ulp on about a third of
+/// inputs. That is invisible after the 4-decimal rounding on every
+/// reference fixture, and far below the engine's own distance from the
+/// reference's logits, so it is not reproduced; the steps either side of
+/// the `exp` are.
+fn calibrate(logits: &[f32], t: f64) -> Vec<f32> {
+    let z = scale(logits, t);
+    let max = z.iter().copied().fold(f32::NEG_INFINITY, f32::max);
+    let e: Vec<f32> = z.iter().map(|&v| (v - max).exp()).collect();
+    normalize(&e)
+}
+
+/// `logits / t` in float32: numpy casts the Python-float temperature to the
+/// array's dtype rather than promoting the array.
+fn scale(logits: &[f32], t: f64) -> Vec<f32> {
+    let t = t as f32;
+    logits.iter().map(|&l| l / t).collect()
+}
+
+/// `e / e.sum()` with numpy's pairwise float32 sum.
+fn normalize(e: &[f32]) -> Vec<f32> {
+    let sum = pairwise_sum(e);
+    e.iter().map(|&v| v / sum).collect()
+}
+
+/// `1 - H(p) / log(k)`, clipped to [0, 1]: how concentrated the whole
+/// distribution is. Not calibrated; see `answer_confidence` for that.
+fn entropy_confidence(p: &[f32]) -> f32 {
+    let k = p.len();
+    if k < 2 {
+        return 1.0;
+    }
+    let terms: Vec<f32> = p.iter().map(|&v| v * v.clamp(1e-12, 1.0).ln()).collect();
+    let ent = -pairwise_sum(&terms);
+    (1.0 - ent / (k as f64).ln() as f32).clamp(0.0, 1.0)
+}
+
+/// The first index of the largest value, as `argmax` picks it.
+fn argmax(p: &[f32]) -> usize {
+    let mut best = 0;
+    for (i, &v) in p.iter().enumerate() {
+        if v > p[best] {
+            best = i;
+        }
+    }
+    best
+}
+
+/// Python's `round(x, 4)`: correctly rounded, ties to even on the exact
+/// binary value. Rust's fixed-precision formatting rounds the same way.
+fn round4(x: f64) -> f64 {
+    format!("{x:.4}").parse().expect("formatted float parses")
+}
+
+/// numpy's pairwise summation over a contiguous array, which the rounded
+/// outputs depend on: a plain loop below 8 elements, eight interleaved
+/// accumulators up to 128, and recursive halving (at a multiple of 8)
+/// beyond.
+fn pairwise_sum<T>(a: &[T]) -> T
+where
+    T: Copy + std::ops::Add<Output = T> + Default,
+{
+    const BLOCK: usize = 128;
+    let n = a.len();
+    if n < 8 {
+        let mut res = T::default();
+        for &v in a {
+            res = res + v;
+        }
+        res
+    } else if n <= BLOCK {
+        let mut r = [a[0], a[1], a[2], a[3], a[4], a[5], a[6], a[7]];
+        let mut i = 8;
+        while i < n - n % 8 {
+            for j in 0..8 {
+                r[j] = r[j] + a[i + j];
+            }
+            i += 8;
+        }
+        let mut res = ((r[0] + r[1]) + (r[2] + r[3])) + ((r[4] + r[5]) + (r[6] + r[7]));
+        while i < n {
+            res = res + a[i];
+            i += 1;
+        }
+        res
+    } else {
+        let mut n2 = n / 2;
+        n2 -= n2 % 8;
+        pairwise_sum(&a[..n2]) + pairwise_sum(&a[n2..])
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn pairwise_sum_differs_from_a_running_sum_where_numpy_does() {
+        // 1e8 swamps every 1.0 added to it one at a time in f32; pairwise
+        // grouping keeps them. `int(np.float32 array.sum())` of this array is 100000008.
+        let mut a = vec![1.0f32; 16];
+        a[0] = 1e8;
+        let running = a.iter().fold(0.0f32, |s, &v| s + v);
+        assert_eq!(running, 1e8);
+        assert_eq!(pairwise_sum(&a), 100_000_008.0);
+    }
+
+    /// Rows of `testdata/laya/numpy_softmax.txt`: the reference's
+    /// calibration run in numpy on random float32 logits, with the bits of
+    /// every intermediate. Bit equality pins the evaluation order — a
+    /// running sum instead of numpy's pairwise one, or dividing by the
+    /// temperature in float64, changes low bits that the 4-decimal rounding
+    /// usually hides.
+    fn numpy_rows() -> Vec<(f64, [Vec<f32>; 4])> {
+        include_str!("../testdata/laya/numpy_softmax.txt")
+            .lines()
+            .filter(|l| !l.starts_with('#') && !l.trim().is_empty())
+            .map(|l| {
+                let mut f = l.split('|');
+                let t: f64 = f.next().unwrap().parse().unwrap();
+                let bits = |s: &str| -> Vec<f32> {
+                    s.split(',')
+                        .map(|h| f32::from_bits(u32::from_str_radix(&h[2..], 16).unwrap()))
+                        .collect()
+                };
+                let v: Vec<Vec<f32>> = f.map(bits).collect();
+                (t, [v[0].clone(), v[1].clone(), v[2].clone(), v[3].clone()])
+            })
+            .collect()
+    }
+
+    fn bits(v: &[f32]) -> Vec<u32> {
+        v.iter().map(|x| x.to_bits()).collect()
+    }
+
+    #[test]
+    fn calibration_matches_numpy_either_side_of_exp() {
+        let rows = numpy_rows();
+        assert_eq!(rows.len(), 4);
+        for (t, [logits, z, e, p]) in rows {
+            assert_eq!(
+                bits(&scale(&logits, t)),
+                bits(&z),
+                "k={} t={t}: z",
+                logits.len()
+            );
+            assert_eq!(
+                bits(&normalize(&e)),
+                bits(&p),
+                "k={} t={t}: p",
+                logits.len()
+            );
+        }
+    }
+
+    #[test]
+    fn round4_ties_to_even_on_the_binary_value() {
+        assert_eq!(round4(0.03125), 0.0312);
+        assert_eq!(round4(1.00005), 1.0001); // 1.00005 is just above the tie in binary
+        assert_eq!(round4(0.99995), 1.0);
+    }
+}

crates/neuron/src/harness/decision/tokenizer.rs

@@ -0,0 +1,81 @@
+//! The tokenizer a decision checkpoint ships, and its special tokens.
+
+use std::path::Path;
+
+/// The special tokens the sequence layout uses, resolved to ids.
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct SpecialTokens {
+    /// The option marker's text (`[MASK]`, `<mask>` for mmBERT). Literal
+    /// occurrences in caller text are replaced with a space before
+    /// tokenizing, so no caller can forge a marker.
+    pub mask_token: String,
+    pub cls_id: u32,
+    pub sep_id: u32,
+    pub mask_id: u32,
+    /// Padding for batched rows; not part of any sequence.
+    pub pad_id: u32,
+}
+
+/// Text → token ids, with no special tokens added.
+pub trait TextEncoder {
+    fn encode(&self, text: &str) -> Result<Vec<u32>, String>;
+    fn special(&self) -> &SpecialTokens;
+}
+
+/// A checkpoint's `tokenizer/` directory: `tokenizer.json` plus the special
+/// token names in `tokenizer_config.json`.
+pub struct DecisionTokenizer {
+    tok: tokenizers::Tokenizer,
+    special: SpecialTokens,
+}
+
+impl DecisionTokenizer {
+    pub fn from_dir(dir: &Path) -> anyhow::Result<Self> {
+        let mut tok = tokenizers::Tokenizer::from_file(dir.join("tokenizer.json"))
+            .map_err(|e| anyhow::anyhow!("{}: {e}", dir.join("tokenizer.json").display()))?;
+        // The Python tokenizer call the reference makes pads and truncates
+        // nothing unless asked; a tokenizer.json that carries either setting
+        // would otherwise apply it here.
+        tok.with_truncation(None)
+            .map_err(|e| anyhow::anyhow!("disabling truncation: {e}"))?;
+        tok.with_padding(None);
+        let cfg_path = dir.join("tokenizer_config.json");
+        let cfg: serde_json::Value = serde_json::from_str(&std::fs::read_to_string(&cfg_path)?)
+            .map_err(|e| anyhow::anyhow!("{}: {e}", cfg_path.display()))?;
+        let name = |key: &str| -> anyhow::Result<String> {
+            cfg.get(key)
+                .and_then(|v| {
+                    v.as_str()
+                        .or_else(|| v.get("content").and_then(|c| c.as_str()))
+                })
+                .map(str::to_owned)
+                .ok_or_else(|| anyhow::anyhow!("{}: no {key}", cfg_path.display()))
+        };
+        let id = |token: &str| -> anyhow::Result<u32> {
+            tok.token_to_id(token)
+                .ok_or_else(|| anyhow::anyhow!("special token {token:?} is not in the vocabulary"))
+        };
+        let mask_token = name("mask_token")?;
+        let special = SpecialTokens {
+            cls_id: id(&name("cls_token")?)?,
+            sep_id: id(&name("sep_token")?)?,
+            mask_id: id(&mask_token)?,
+            pad_id: id(&name("pad_token")?)?,
+            mask_token,
+        };
+        Ok(Self { tok, special })
+    }
+}
+
+impl TextEncoder for DecisionTokenizer {
+    fn encode(&self, text: &str) -> Result<Vec<u32>, String> {
+        self.tok
+            .encode(text, false)
+            .map(|e| e.get_ids().to_vec())
+            .map_err(|e| e.to_string())
+    }
+
+    fn special(&self) -> &SpecialTokens {
+        &self.special
+    }
+}

crates/neuron/src/harness/decision_model.rs

@@ -0,0 +1,252 @@
+//! A loaded Laya decision model, as the rest of neuron sees it (#336).
+//!
+//! The weights live in a device worker's decision slab (CUDA) or in an
+//! `Arc` scored on the blocking pool (CPU); either way the caller holds
+//! only this struct and asks it to [`DecisionCheckpoint::decide`] a
+//! batch of already-assembled questions. Turning a `/v1/systemone`
+//! request into that batch — rendering options, budgets, tokenisation —
+//! and turning the scores back into calibrated answers happen outside
+//! this module.
+//!
+//! One Hugging Face repo can carry several checkpoints (the released
+//! `convaiinnovations/laya` has the English one at its root and
+//! `multilingual/` and `typed-decisions/` beside it), so a loaded model
+//! is a list of checkpoints, each named by its subfolder.
+
+use std::path::PathBuf;
+use std::sync::Arc;
+use std::sync::atomic::AtomicBool;
+
+use anyhow::{Context, Result};
+use candle_core::DType;
+
+use super::arch::laya::{CheckpointConfig, DecisionBatch, DecisionRow, LayaModel};
+use super::decision::{DecisionConfig, DecisionTokenizer};
+use super::device_worker::{DecisionHandle, DeviceWorkerHandle};
+use cortex_core::harness::ModelSpec;
+
+/// Routing name of the checkpoint at the repo root. A released Laya
+/// family's root is its English checkpoint, and upstream routes to it
+/// under that name.
+pub const ROOT_CHECKPOINT: &str = super::decision::route::ENGLISH;
+
+/// Where one checkpoint's weights live and how to reach them.
+enum Backend {
+    /// Owned by a CUDA device worker; scored through its channel.
+    Worker {
+        worker: Arc<DeviceWorkerHandle>,
+        handle: DecisionHandle,
+    },
+    /// CPU fallback. `LayaModel::forward` takes `&self`, so concurrent
+    /// requests share it without a lock.
+    Cpu(Arc<LayaModel>),
+}
+
+/// The files one checkpoint needs, already on local disk.
+#[derive(Debug, Clone)]
+pub struct CheckpointFiles {
+    /// The checkpoint directory: holds `rl_agent_config.json`,
+    /// `encoder/config.json` and `model.safetensors`.
+    pub dir: PathBuf,
+    pub tokenizer: PathBuf,
+}
+
+pub struct DecisionCheckpoint {
+    /// The checkpoint's routing name: [`ROOT_CHECKPOINT`] for the repo
+    /// root, else its subfolder (`multilingual`, `typed-decisions`).
+    pub name: String,
+    pub config: CheckpointConfig,
+    /// Budgets and fitted temperatures from `rl_agent_config.json`, as
+    /// the prompt builder and calibration read them.
+    pub calibration: DecisionConfig,
+    pub tokenizer: DecisionTokenizer,
+    /// Precision the weights were loaded in.
+    pub dtype: DType,
+    backend: Backend,
+}
+
+impl DecisionCheckpoint {
+    /// Load one checkpoint onto `worker`'s device, or onto the CPU when
+    /// `worker` is `None`.
+    pub async fn load(
+        name: String,
+        files: CheckpointFiles,
+        model_id: &str,
+        worker: Option<Arc<DeviceWorkerHandle>>,
+    ) -> Result<Self> {
+        let config = CheckpointConfig::from_dir(&files.dir)?;
+        let calibration_path = files.dir.join("rl_agent_config.json");
+        let calibration = DecisionConfig::from_json(
+            &std::fs::read_to_string(&calibration_path)
+                .with_context(|| format!("read {}", calibration_path.display()))?,
+        )
+        .map_err(|e| anyhow::anyhow!("{e}"))?;
+        let tokenizer_dir = files
+            .tokenizer
+            .parent()
+            .context("tokenizer.json has no parent directory")?;
+        let tokenizer = DecisionTokenizer::from_dir(tokenizer_dir)?;
+        let (dtype, backend) = match worker {
+            Some(worker) => {
+                let wanted = device_dtype(config.laya.amp_dtype.as_deref());
+                let (handle, dtype) = worker
+                    .load_decision(files.dir.clone(), model_id.to_string(), wanted)
+                    .await
+                    .map_err(|e| anyhow::anyhow!("worker load_decision: {e:#}"))?;
+                (dtype, Backend::Worker { worker, handle })
+            }
+            None => {
+                let dir = files.dir.clone();
+                let model = tokio::task::spawn_blocking(move || {
+                    LayaModel::load(&dir, &candle_core::Device::Cpu, DType::F32)
+                })
+                .await
+                .context("decision load task")??;
+                (DType::F32, Backend::Cpu(Arc::new(model)))
+            }
+        };
+        Ok(Self {
+            name,
+            config,
+            calibration,
+            tokenizer,
+            dtype,
+            backend,
+        })
+    }
+
+    /// Score one batch: one [`DecisionRow`] per sequence, in order.
+    pub async fn decide(&self, batch: DecisionBatch) -> Result<Vec<DecisionRow>> {
+        match &self.backend {
+            Backend::Worker { worker, handle } => worker
+                .decide(*handle, batch)
+                .await
+                .map_err(|e| anyhow::anyhow!("{e:#}")),
+            Backend::Cpu(model) => {
+                let model = Arc::clone(model);
+                tokio::task::spawn_blocking(move || model.forward(&batch))
+                    .await
+                    .context("decision forward task")?
+            }
+        }
+    }
+
+    /// Release the weights. On a worker this drops them on the thread
+    /// that owns the device context.
+    pub async fn release(&self) {
+        if let Backend::Worker { worker, handle } = &self.backend
+            && let Err(e) = worker.drop_decision(*handle).await
+        {
+            tracing::warn!(
+                checkpoint = %self.name,
+                error = %e,
+                "decision unload: DropDecision RPC failed (weights may leak in worker slab)"
+            );
+        }
+    }
+}
+
+/// The precision to serve in on a GPU: the autocast dtype the reference
+/// itself runs there (`amp_dtype` in `rl_agent_config.json`), else bf16.
+fn device_dtype(amp_dtype: Option<&str>) -> DType {
+    match amp_dtype {
+        Some("fp16") | Some("float16") => DType::F16,
+        Some("fp32") | Some("float32") => DType::F32,
+        _ => DType::BF16,
+    }
+}
+
+pub struct LoadedDecisionModel {
+    pub model_id: String,
+    pub spec: ModelSpec,
+    pub devices: Vec<u32>,
+    /// At least one; the first is the default.
+    pub checkpoints: Vec<DecisionCheckpoint>,
+    /// Set when a forward returns a driver-level error, mirroring the
+    /// other model kinds: `/models` reports it and requests fast-reject.
+    pub poisoned: Arc<AtomicBool>,
+    pub admission: super::admission::AdmissionController,
+}
+
+impl LoadedDecisionModel {
+    /// The checkpoint called `name`, or the default one for `None`.
+    pub fn checkpoint(&self, name: Option<&str>) -> Option<&DecisionCheckpoint> {
+        match name {
+            None => self.checkpoints.first(),
+            Some(n) => self.checkpoints.iter().find(|c| c.name == n),
+        }
+    }
+
+    pub async fn release(&self) {
+        for c in &self.checkpoints {
+            c.release().await;
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn tiny() -> CheckpointFiles {
+        let dir =
+            std::path::Path::new(env!("CARGO_MANIFEST_DIR")).join("src/harness/testdata/laya/tiny");
+        CheckpointFiles {
+            tokenizer: dir.join("tokenizer.json"),
+            dir,
+        }
+    }
+
+    #[test]
+    fn gpu_dtype_follows_the_reference_autocast() {
+        assert_eq!(device_dtype(Some("bf16")), DType::BF16);
+        assert_eq!(device_dtype(Some("fp16")), DType::F16);
+        assert_eq!(device_dtype(None), DType::BF16);
+    }
+
+    /// The whole async path on the CPU backend, and the device-worker
+    /// path on a CPU build (the worker runs, with a CPU device), must
+    /// give the same scores as the model called directly.
+    #[tokio::test]
+    async fn both_backends_score_like_the_model() {
+        let files = tiny();
+        let direct = LayaModel::load(&files.dir, &candle_core::Device::Cpu, DType::F32).unwrap();
+        let batch = DecisionBatch {
+            seqs: vec![super::super::arch::laya::DecisionSeq {
+                input_ids: vec![5, 9, 11, 3, 7, 2],
+                markers: vec![1, 3],
+                qtype: 0,
+            }],
+        };
+        let want = direct.forward(&batch).unwrap();
+
+        let cpu = DecisionCheckpoint::load("root".into(), files.clone(), "tiny", None)
+            .await
+            .unwrap();
+        assert_eq!(cpu.decide(batch.clone()).await.unwrap(), want);
+
+        // A CPU build's worker holds a CPU device, so this exercises the
+        // job round-trip (load, decide, drop) without a GPU — and the
+        // worker must fall back to f32 there, reporting that it did.
+        let worker = DeviceWorkerHandle::spawn(0).unwrap();
+        let on_worker =
+            DecisionCheckpoint::load("root".into(), files, "tiny", Some(worker.clone()))
+                .await
+                .unwrap();
+        #[cfg(not(feature = "cuda"))]
+        {
+            assert_eq!(on_worker.dtype, DType::F32);
+            assert_eq!(on_worker.decide(batch).await.unwrap(), want);
+        }
+        #[cfg(feature = "cuda")]
+        {
+            let got = on_worker.decide(batch).await.unwrap();
+            for (g, w) in got[0].logits.iter().zip(&want[0].logits) {
+                assert!((g - w).abs() < 5e-2, "{g} vs {w}");
+            }
+        }
+        on_worker.release().await;
+        assert!(on_worker.decide(DecisionBatch::default()).await.is_err());
+        worker.shutdown().unwrap();
+    }
+}

crates/neuron/src/harness/device_worker/dispatch.rs

@@ -13,12 +13,13 @@
 //! ARCH model state in this state slab will gain a companion
 //! `tp_models: HashMap<TpHandle, Box<TpLeaderModel>>`.
 
+use crate::harness::arch::laya::LayaModel;
 use crate::harness::arch::snapshot::KvCacheSnapshot;
 use crate::harness::candle::ModelArch;
 #[cfg(feature = "cuda")]
 use crate::harness::device_worker::jobs::TpHandle;
 use crate::harness::device_worker::jobs::{
-    ArchHandle, DenseLoad, ImageHandle, ImageInput, Job, KvSnapshotId,
+    ArchHandle, DecisionHandle, DenseLoad, ImageHandle, ImageInput, Job, KvSnapshotId,
 };
 use crate::harness::image::ZImagePipeline;
 #[cfg(feature = "cuda")]
@@ -65,6 +66,11 @@ struct DeviceWorkerState {
     image_models: HashMap<ImageHandle, Box<ZImagePipeline>>,
     /// Counter for minting fresh `ImageHandle`s.
     next_image_handle: u64,
+    /// Laya decision-model slab (#336). Same lifecycle discipline as
+    /// `image_models`, in its own namespace.
+    decision_models: HashMap<DecisionHandle, Box<LayaModel>>,
+    /// Counter for minting fresh `DecisionHandle`s.
+    next_decision_handle: u64,
     /// Leader's NCCL state. Populated by `Job::NcclInit`; the
     /// underlying `Comm`'s libnccl handle lives bound to this thread
     /// for its entire lifetime. Subprocess workers maintain their own
@@ -322,6 +328,15 @@ pub(crate) fn run(device_index: u32, rx: Receiver<Job>, poisoned: Arc<AtomicBool
                 }
                 let _ = reply.send(result);
             }
+            Job::ForwardLogitsMulti {
+                handle,
+                tokens,
+                offset,
+                reply,
+            } => {
+                let result = forward_logits_multi(&mut state, handle, &tokens, offset);
+                let _ = reply.send(result);
+            }
             Job::ForwardLogitsBatch {
                 handle,
                 tokens,
@@ -599,6 +614,61 @@ pub(crate) fn run(device_index: u32, rx: Receiver<Job>, poisoned: Arc<AtomicBool
                 let _ = reply.send(result);
             }
             #[cfg(feature = "cuda")]
+            Job::TpLoadMtpHead {
+                handle,
+                config_json,
+                safetensors_paths,
+                reply,
+            } => {
+                let result = tp_load_mtp_head(&mut state, handle, &config_json, &safetensors_paths);
+                let _ = reply.send(result);
+            }
+            #[cfg(feature = "cuda")]
+            Job::TpMtpPrefillChunk {
+                handle,
+                shifted,
+                start_pos,
+                reply,
+            } => {
+                let result = tp_mtp(&mut state, handle, |m| {
+                    m.mtp_prefill_chunk(&shifted, start_pos)
+                });
+                let _ = reply.send(result);
+            }
+            #[cfg(feature = "cuda")]
+            Job::TpMtpDraft {
+                handle,
+                first_token,
+                start_pos,
+                k,
+                reply,
+            } => {
+                let result = tp_mtp(&mut state, handle, |m| {
+                    m.mtp_draft(first_token, start_pos, k)
+                });
+                let _ = reply.send(result);
+            }
+            #[cfg(feature = "cuda")]
+            Job::TpMtpAdvance {
+                handle,
+                token,
+                pos,
+                reply,
+            } => {
+                let result = tp_mtp(&mut state, handle, |m| m.mtp_advance(token, pos));
+                let _ = reply.send(result);
+            }
+            #[cfg(feature = "cuda")]
+            Job::TpForwardLogitsMulti {
+                handle,
+                tokens,
+                offset,
+                reply,
+            } => {
+                let result = tp_forward_logits_multi(&mut state, handle, &tokens, offset);
+                let _ = reply.send(result);
+            }
+            #[cfg(feature = "cuda")]
             Job::TpForwardLogitsWithImages {
                 handle,
                 tokens,
@@ -687,6 +757,62 @@ pub(crate) fn run(device_index: u32, rx: Receiver<Job>, poisoned: Arc<AtomicBool
                 trim_device_pool(&state);
                 let _ = reply.send(result);
             }
+            Job::LoadDecision {
+                dir,
+                model_id,
+                dtype,
+                reply,
+            } => {
+                let dtype = if state.device.is_cpu() {
+                    candle_core::DType::F32
+                } else {
+                    dtype
+                };
+                let result = LayaModel::load(&dir, &state.device, dtype)
+                    .with_context(|| format!("load laya checkpoint {}", dir.display()))
+                    .map(|model| {
+                        let handle = DecisionHandle(state.next_decision_handle);
+                        state.next_decision_handle = state.next_decision_handle.wrapping_add(1);
+                        state.decision_models.insert(handle, Box::new(model));
+                        tracing::info!(
+                            device_index,
+                            model_id,
+                            checkpoint = %dir.display(),
+                            handle = handle.0,
+                            "device worker: decision model loaded"
+                        );
+                        (handle, dtype)
+                    });
+                trim_device_pool(&state);
+                let _ = reply.send(result);
+            }
+            Job::DropDecision { handle, reply } => {
+                let removed = state.decision_models.remove(&handle);
+                let was_present = removed.is_some();
+                // Explicit drop on this thread, as for DropArch/DropImage.
+                drop(removed);
+                if was_present {
+                    trim_device_pool(&state);
+                }
+                tracing::debug!(
+                    device_index,
+                    handle = handle.0,
+                    was_present,
+                    "device worker: decision model dropped"
+                );
+                let _ = reply.send(());
+            }
+            Job::Decide {
+                handle,
+                batch,
+                reply,
+            } => {
+                let result = match state.decision_models.get(&handle) {
+                    Some(model) => model.forward(&batch),
+                    None => Err(anyhow::anyhow!("no decision model for handle {}", handle.0)),
+                };
+                let _ = reply.send(result);
+            }
             Job::Shutdown => unreachable!("Shutdown should break above"),
         }
     }
@@ -759,6 +885,8 @@ fn init_state(device_index: u32) -> DeviceWorkerState {
             next_kv_snapshot_id: 1,
             image_models: HashMap::new(),
             next_image_handle: 1,
+            decision_models: HashMap::new(),
+            next_decision_handle: 1,
             nccl: NcclState::new(),
             tp_models: HashMap::new(),
             next_tp_handle: 1,
@@ -777,6 +905,8 @@ fn init_state(device_index: u32) -> DeviceWorkerState {
             next_kv_snapshot_id: 1,
             image_models: HashMap::new(),
             next_image_handle: 1,
+            decision_models: HashMap::new(),
+            next_decision_handle: 1,
             nccl: NcclState::new(),
         }
     }
@@ -1305,6 +1435,85 @@ fn tp_forward_logits(
     Ok(values)
 }
 
+/// Run something on the leader's TP draft head (#96).
+///
+/// The head only exists on the `qwen3_5` arch and only on a model that
+/// was asked to load one, so both misses are reported rather than
+/// silently skipped.
+#[cfg(feature = "cuda")]
+fn tp_mtp<T>(
+    state: &mut DeviceWorkerState,
+    handle: TpHandle,
+    f: impl FnOnce(&mut super::super::tp::tp_qwen3_5::TpQwen3_5ForCausalLM) -> anyhow::Result<T>,
+) -> anyhow::Result<T> {
+    let model = state
+        .tp_models
+        .get_mut(&handle)
+        .ok_or_else(|| anyhow::anyhow!("tp mtp: no model for handle {}", handle.0))?;
+    match &mut **model {
+        super::super::tp::TpLeaderModel::Qwen3_5(m) => f(m),
+        super::super::tp::TpLeaderModel::Qwen3(_) => {
+            anyhow::bail!("tp mtp: the qwen3 arch ships no MTP head")
+        }
+    }
+}
+
+/// Load the draft head onto the leader's TP model (#96).
+#[cfg(feature = "cuda")]
+fn tp_load_mtp_head(
+    state: &mut DeviceWorkerState,
+    handle: TpHandle,
+    config_json: &str,
+    safetensors_paths: &[String],
+) -> anyhow::Result<()> {
+    use candle_core::DType;
+
+    let cfg = super::super::tp::tp_qwen3_5::Config::from_config_json(config_json)
+        .context("tp mtp: parse config.json")?;
+    let paths: Vec<std::path::PathBuf> = safetensors_paths.iter().map(Into::into).collect();
+    let device = state.device.clone();
+    // SAFETY: the same checkpoint files the shard load already mmapped.
+    let vb = unsafe {
+        candle_nn::var_builder::ShardedSafeTensors::var_builder(&paths, DType::BF16, &device)?
+    };
+    tp_mtp(state, handle, |m| m.load_mtp_head(&cfg.text_config, &vb))
+}
+
+/// TP mirror of [`forward_logits_multi`] on the leader's shard (#96).
+/// The caller has already fanned the matching `GenerateStepMulti` out
+/// to the subprocess ranks; the `AllReduce` collectives inside this
+/// forward block until they arrive.
+#[cfg(feature = "cuda")]
+fn tp_forward_logits_multi(
+    state: &mut DeviceWorkerState,
+    handle: TpHandle,
+    tokens: &[u32],
+    offset: usize,
+) -> anyhow::Result<Vec<Vec<f32>>> {
+    use candle_core::{DType, Tensor};
+
+    anyhow::ensure!(
+        !tokens.is_empty(),
+        "TpForwardLogitsMulti: empty token slice"
+    );
+    let input = Tensor::new(tokens, &state.device)?.unsqueeze(0)?;
+    let model = state
+        .tp_models
+        .get_mut(&handle)
+        .ok_or_else(|| anyhow::anyhow!("TpForwardLogitsMulti: no model for handle {}", handle.0))?;
+
+    // (1, L, vocab) → one CPU row per position.
+    let logits = model.forward_multi(&input, offset)?.to_dtype(DType::F32)?;
+    let (_, positions, vocab) = logits.dims3()?;
+    anyhow::ensure!(
+        positions == tokens.len(),
+        "TpForwardLogitsMulti: model returned {positions} rows for {} tokens",
+        tokens.len()
+    );
+    let flat: Vec<f32> = logits.flatten_all()?.to_vec1()?;
+    Ok(flat.chunks(vocab).map(<[f32]>::to_vec).collect())
+}
+
 /// TP-equivalent of [`forward_logits_batch`] on the leader's shard
 /// (#98). The caller has already fanned the matching
 /// `GenerateStepBatch` out to the subprocess ranks.
@@ -1575,6 +1784,40 @@ fn extract_kv_rows(
 /// on the worker's device, derives per-row positions and the padding
 /// mask, and copies each row's logits back to CPU — same
 /// "tensors never escape the worker" contract as `forward_logits`.
+/// Logits at every position of a multi-token forward (#96).
+///
+/// The tensors stay on the worker thread; only CPU rows cross back, one
+/// per position. At K+1 = 5 positions and a 248k vocabulary that is
+/// ~5 MB a round — acceptable while this is a measurement path, and the
+/// first thing to replace with a device-side argmax if the draft/verify
+/// loop ever runs on it in production.
+fn forward_logits_multi(
+    state: &mut DeviceWorkerState,
+    handle: ArchHandle,
+    tokens: &[u32],
+    offset: usize,
+) -> anyhow::Result<Vec<Vec<f32>>> {
+    use candle_core::{DType, Tensor};
+
+    anyhow::ensure!(!tokens.is_empty(), "ForwardLogitsMulti: empty token slice");
+    let input = Tensor::new(tokens, &state.device)?.unsqueeze(0)?;
+    let arch = state
+        .models
+        .get_mut(&handle)
+        .ok_or_else(|| anyhow::anyhow!("ForwardLogitsMulti: no model for handle {}", handle.0))?;
+
+    // (1, L, vocab) → one CPU row per position.
+    let logits = arch.forward_multi(&input, offset)?.to_dtype(DType::F32)?;
+    let (_, positions, vocab) = logits.dims3()?;
+    anyhow::ensure!(
+        positions == tokens.len(),
+        "ForwardLogitsMulti: model returned {positions} rows for {} tokens",
+        tokens.len()
+    );
+    let flat: Vec<f32> = logits.flatten_all()?.to_vec1()?;
+    Ok(flat.chunks(vocab).map(<[f32]>::to_vec).collect())
+}
+
 fn forward_logits_batch(
     state: &mut DeviceWorkerState,
     handle: ArchHandle,
@@ -1771,6 +2014,9 @@ fn drain_poisoned(job: Job, device_index: u32) {
         Job::AssembleKvBatch { reply, .. } => {
             let _ = reply.send(Err(err()));
         }
+        Job::ForwardLogitsMulti { reply, .. } => {
+            let _ = reply.send(Err(err()));
+        }
         Job::ForwardLogitsBatch { reply, .. } => {
             let _ = reply.send(Err(err()));
         }
@@ -1789,6 +2035,15 @@ fn drain_poisoned(job: Job, device_index: u32) {
         Job::GenerateImage { reply, .. } => {
             let _ = reply.send(Err(err()));
         }
+        Job::LoadDecision { reply, .. } => {
+            let _ = reply.send(Err(err()));
+        }
+        Job::DropDecision { reply, .. } => {
+            let _ = reply.send(());
+        }
+        Job::Decide { reply, .. } => {
+            let _ = reply.send(Err(err()));
+        }
         Job::ForwardLogitsWithImages { reply, .. } => {
             let _ = reply.send(Err(err()));
         }
@@ -1847,6 +2102,26 @@ fn drain_poisoned(job: Job, device_index: u32) {
             let _ = reply.send(());
         }
         #[cfg(feature = "cuda")]
+        Job::TpLoadMtpHead { reply, .. } => {
+            let _ = reply.send(Err(err()));
+        }
+        #[cfg(feature = "cuda")]
+        Job::TpMtpPrefillChunk { reply, .. } => {
+            let _ = reply.send(Err(err()));
+        }
+        #[cfg(feature = "cuda")]
+        Job::TpMtpDraft { reply, .. } => {
+            let _ = reply.send(Err(err()));
+        }
+        #[cfg(feature = "cuda")]
+        Job::TpMtpAdvance { reply, .. } => {
+            let _ = reply.send(Err(err()));
+        }
+        #[cfg(feature = "cuda")]
+        Job::TpForwardLogitsMulti { reply, .. } => {
+            let _ = reply.send(Err(err()));
+        }
+        #[cfg(feature = "cuda")]
         Job::TpForwardLogits { reply, .. } => {
             let _ = reply.send(Err(err()));
         }

crates/neuron/src/harness/device_worker/jobs.rs

@@ -59,6 +59,15 @@ pub struct TpHandle(pub u64);
 #[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
 pub struct ImageHandle(pub u64);
 
+/// Opaque handle to a Laya decision model ([`LayaModel`]) stored in the
+/// worker thread's decision slab (#336). Same discipline as
+/// [`ImageHandle`]: construction, every forward and `Drop` happen on the
+/// worker thread, and `Job::DropDecision` is the only way to release it.
+///
+/// [`LayaModel`]: crate::harness::arch::laya::LayaModel
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
+pub struct DecisionHandle(pub u64);
+
 /// Opaque handle to a prefix-cache snapshot (#11) stored worker-side
 /// next to the model slab. Scoped to the `ArchHandle` it was captured
 /// from — `Job::DropArch` drops every snapshot under its handle. The
@@ -225,6 +234,19 @@ pub enum Job {
     /// values `AssembleKvBatch` was built from) and replies one CPU
     /// `[vocab]` logits row per batch row, ready for per-slot
     /// sampling on the async side.
+    /// Logits at every position of a multi-token forward (#96).
+    ///
+    /// `ForwardLogits` returns one `[vocab]` row because a decode step
+    /// only has one. A speculative verify pass forwards K+1 tokens and
+    /// needs the target's token at each of them, so this returns one
+    /// row per position — `Vec<Vec<f32>>`, like `ForwardLogitsBatch`,
+    /// but indexed by position rather than by batch row.
+    ForwardLogitsMulti {
+        handle: ArchHandle,
+        tokens: Vec<u32>,
+        offset: usize,
+        reply: oneshot::Sender<Result<Vec<Vec<f32>>>>,
+    },
     ForwardLogitsBatch {
         handle: ArchHandle,
         tokens: Vec<u32>,
@@ -381,6 +403,59 @@ pub enum Job {
     /// without holding a device tensor. The caller is also
     /// responsible for fan-out to subprocess ranks and drain — only
     /// the leader's forward moves into the worker thread.
+    /// Load the MTP draft head onto the leader's TP model (#96).
+    ///
+    /// Leader-only and deliberately not part of `TpLoadShard`: the
+    /// subprocess ranks have no use for a head they never run, and an
+    /// unsharded head on every rank would cost 810 MB each to sit idle.
+    #[cfg(feature = "cuda")]
+    TpLoadMtpHead {
+        handle: TpHandle,
+        config_json: String,
+        safetensors_paths: Vec<String>,
+        reply: oneshot::Sender<Result<()>>,
+    },
+    /// Walk the leader's draft head over a prompt chunk (#96), using
+    /// the hidden state the target's forward just produced.
+    #[cfg(feature = "cuda")]
+    TpMtpPrefillChunk {
+        handle: TpHandle,
+        /// The chunk's tokens shifted one left — the head trails the
+        /// target by a token.
+        shifted: Vec<u32>,
+        start_pos: usize,
+        reply: oneshot::Sender<Result<()>>,
+    },
+    /// Draft `k` tokens on the leader and put the head's cache back
+    /// (#96). Commits nothing: the snapshot/restore pair lives inside
+    /// the job so a caller cannot forget it.
+    #[cfg(feature = "cuda")]
+    TpMtpDraft {
+        handle: TpHandle,
+        first_token: u32,
+        start_pos: usize,
+        k: usize,
+        reply: oneshot::Sender<Result<Vec<u32>>>,
+    },
+    /// Advance the leader's draft head over one committed token (#96).
+    #[cfg(feature = "cuda")]
+    TpMtpAdvance {
+        handle: TpHandle,
+        token: u32,
+        pos: usize,
+        reply: oneshot::Sender<Result<()>>,
+    },
+    /// TP mirror of `ForwardLogitsMulti` (#96): logits at every
+    /// position of a multi-token forward on the leader's shard. The
+    /// caller has already fanned the matching `GenerateStepMulti` out
+    /// to the subprocess ranks, whose collectives this one waits on.
+    #[cfg(feature = "cuda")]
+    TpForwardLogitsMulti {
+        handle: TpHandle,
+        tokens: Vec<u32>,
+        offset: usize,
+        reply: oneshot::Sender<Result<Vec<Vec<f32>>>>,
+    },
     #[cfg(feature = "cuda")]
     TpForwardLogits {
         handle: TpHandle,
@@ -483,6 +558,33 @@ pub enum Job {
         max_dim: usize,
         reply: oneshot::Sender<Result<crate::harness::image::ImageGenResult>>,
     },
+    /// Load one Laya checkpoint directory (#336) — the repo root or one
+    /// of its subfolders, already on local disk — into the decision
+    /// slab. File resolution (hf-hub) happens on the async side.
+    ///
+    /// `dtype` is the precision wanted on a GPU; a worker whose device
+    /// is the CPU loads in f32 instead (the CPU backend has no bf16
+    /// matmul). The reply carries the dtype actually used.
+    LoadDecision {
+        dir: PathBuf,
+        model_id: String,
+        dtype: candle_core::DType,
+        reply: oneshot::Sender<Result<(DecisionHandle, candle_core::DType)>>,
+    },
+    /// Remove a decision model from the slab and drop it on the worker
+    /// thread. Mirrors `DropArch` / `DropImage`.
+    DropDecision {
+        handle: DecisionHandle,
+        reply: oneshot::Sender<()>,
+    },
+    /// Score one batch of assembled questions (#336). One job per
+    /// request: the whole batch is a single encoder forward, and the
+    /// reply is CPU-side scores — no device tensor crosses back.
+    Decide {
+        handle: DecisionHandle,
+        batch: crate::harness::arch::laya::DecisionBatch,
+        reply: oneshot::Sender<Result<Vec<crate::harness::arch::laya::DecisionRow>>>,
+    },
     /// Tell the worker to break its dispatch loop and exit. Any jobs
     /// queued after this in the channel reply `Err` to their oneshot
     /// senders (the senders are dropped on the worker's exit, which

crates/neuron/src/harness/device_worker/mod.rs

@@ -51,7 +51,7 @@ use tokio::sync::oneshot;
 
 #[cfg(feature = "cuda")]
 pub use jobs::TpHandle;
-pub use jobs::{ArchHandle, DenseLoad, ImageHandle, Job, KvSnapshotId};
+pub use jobs::{ArchHandle, DecisionHandle, DenseLoad, ImageHandle, Job, KvSnapshotId};
 
 /// Errors returned by `DeviceWorkerHandle` submit methods.
 #[derive(Debug, thiserror::Error)]
@@ -337,6 +337,85 @@ impl DeviceWorkerHandle {
         }
     }
 
+    /// Load one Laya checkpoint directory (#336) on the worker thread.
+    /// Returns the handle and the dtype the weights were loaded in.
+    pub async fn load_decision(
+        &self,
+        dir: std::path::PathBuf,
+        model_id: String,
+        dtype: candle_core::DType,
+    ) -> Result<(DecisionHandle, candle_core::DType), WorkerError> {
+        if self.poisoned.load(Ordering::Acquire) {
+            return Err(WorkerError::Poisoned {
+                device_index: self.device_index,
+            });
+        }
+        let (reply_tx, reply_rx) = oneshot::channel();
+        self.tx
+            .send(Job::LoadDecision {
+                dir,
+                model_id,
+                dtype,
+                reply: reply_tx,
+            })
+            .map_err(|_| WorkerError::Gone {
+                device_index: self.device_index,
+            })?;
+        match reply_rx.await {
+            Ok(result) => result.map_err(WorkerError::from),
+            Err(_) => Err(WorkerError::Gone {
+                device_index: self.device_index,
+            }),
+        }
+    }
+
+    /// Drop the decision model for `handle` on the worker thread. Same
+    /// poison-tolerant contract as [`Self::drop_arch`].
+    pub async fn drop_decision(&self, handle: DecisionHandle) -> Result<(), WorkerError> {
+        let (reply_tx, reply_rx) = oneshot::channel();
+        self.tx
+            .send(Job::DropDecision {
+                handle,
+                reply: reply_tx,
+            })
+            .map_err(|_| WorkerError::Gone {
+                device_index: self.device_index,
+            })?;
+        reply_rx.await.map_err(|_| WorkerError::Gone {
+            device_index: self.device_index,
+        })
+    }
+
+    /// Score one batch of assembled questions (#336). The reply is
+    /// CPU-side scores; no device tensor crosses back.
+    pub async fn decide(
+        &self,
+        handle: DecisionHandle,
+        batch: crate::harness::arch::laya::DecisionBatch,
+    ) -> Result<Vec<crate::harness::arch::laya::DecisionRow>, WorkerError> {
+        if self.poisoned.load(Ordering::Acquire) {
+            return Err(WorkerError::Poisoned {
+                device_index: self.device_index,
+            });
+        }
+        let (reply_tx, reply_rx) = oneshot::channel();
+        self.tx
+            .send(Job::Decide {
+                handle,
+                batch,
+                reply: reply_tx,
+            })
+            .map_err(|_| WorkerError::Gone {
+                device_index: self.device_index,
+            })?;
+        match reply_rx.await {
+            Ok(result) => result.map_err(WorkerError::from),
+            Err(_) => Err(WorkerError::Gone {
+                device_index: self.device_index,
+            }),
+        }
+    }
+
     /// Tell the worker to drop the `ModelArch` for `handle` on the
     /// worker thread (so CUDA tensors release on the right context).
     /// Returns `Ok(())` even if the handle wasn't in the slab — Drop
@@ -578,6 +657,40 @@ impl DeviceWorkerHandle {
         }
     }
 
+    /// Multi-position forward (#96): logits at every position of
+    /// `tokens`, one CPU `[vocab]` row each. Same no-device-tensor
+    /// contract as [`Self::forward_logits`]; the speculative verify
+    /// pass reads the target's token at each drafted position.
+    pub async fn forward_logits_multi(
+        &self,
+        handle: ArchHandle,
+        tokens: Vec<u32>,
+        offset: usize,
+    ) -> Result<Vec<Vec<f32>>, WorkerError> {
+        if self.poisoned.load(Ordering::Acquire) {
+            return Err(WorkerError::Poisoned {
+                device_index: self.device_index,
+            });
+        }
+        let (reply_tx, reply_rx) = oneshot::channel();
+        self.tx
+            .send(Job::ForwardLogitsMulti {
+                handle,
+                tokens,
+                offset,
+                reply: reply_tx,
+            })
+            .map_err(|_| WorkerError::Gone {
+                device_index: self.device_index,
+            })?;
+        match reply_rx.await {
+            Ok(result) => result.map_err(WorkerError::from),
+            Err(_) => Err(WorkerError::Gone {
+                device_index: self.device_index,
+            }),
+        }
+    }
+
     /// One lockstep batched decode step (#98): row i's next token at
     /// position `prefix_lens[i] + step`. Returns one CPU `[vocab]`
     /// logits row per batch row, ready for per-slot sampling — same
@@ -934,6 +1047,137 @@ impl DeviceWorkerHandle {
     /// logits as `Vec<f32>` ready for sampling. The caller is
     /// responsible for fan-out / drain of the subprocess workers
     /// concurrently with this call.
+    /// Load the MTP draft head onto the leader's TP model (#96).
+    #[cfg(feature = "cuda")]
+    pub async fn tp_load_mtp_head(
+        &self,
+        handle: TpHandle,
+        config_json: String,
+        safetensors_paths: Vec<String>,
+    ) -> Result<(), WorkerError> {
+        self.tp_mtp_job(|reply| Job::TpLoadMtpHead {
+            handle,
+            config_json,
+            safetensors_paths,
+            reply,
+        })
+        .await
+    }
+
+    /// Walk the leader's draft head over a prompt chunk (#96).
+    #[cfg(feature = "cuda")]
+    pub async fn tp_mtp_prefill_chunk(
+        &self,
+        handle: TpHandle,
+        shifted: Vec<u32>,
+        start_pos: usize,
+    ) -> Result<(), WorkerError> {
+        self.tp_mtp_job(|reply| Job::TpMtpPrefillChunk {
+            handle,
+            shifted,
+            start_pos,
+            reply,
+        })
+        .await
+    }
+
+    /// Draft `k` tokens on the leader, rewinding the head afterwards
+    /// (#96). Commits nothing.
+    #[cfg(feature = "cuda")]
+    pub async fn tp_mtp_draft(
+        &self,
+        handle: TpHandle,
+        first_token: u32,
+        start_pos: usize,
+        k: usize,
+    ) -> Result<Vec<u32>, WorkerError> {
+        self.tp_mtp_job(|reply| Job::TpMtpDraft {
+            handle,
+            first_token,
+            start_pos,
+            k,
+            reply,
+        })
+        .await
+    }
+
+    /// Advance the leader's draft head over one committed token (#96).
+    #[cfg(feature = "cuda")]
+    pub async fn tp_mtp_advance(
+        &self,
+        handle: TpHandle,
+        token: u32,
+        pos: usize,
+    ) -> Result<(), WorkerError> {
+        self.tp_mtp_job(|reply| Job::TpMtpAdvance {
+            handle,
+            token,
+            pos,
+            reply,
+        })
+        .await
+    }
+
+    /// Shared submit/await for the leader-only draft-head jobs (#96):
+    /// same poison check and channel handling as every other job, in
+    /// one place because these four differ only in their payload.
+    #[cfg(feature = "cuda")]
+    async fn tp_mtp_job<T>(
+        &self,
+        build: impl FnOnce(oneshot::Sender<anyhow::Result<T>>) -> Job,
+    ) -> Result<T, WorkerError> {
+        if self.poisoned.load(Ordering::Acquire) {
+            return Err(WorkerError::Poisoned {
+                device_index: self.device_index,
+            });
+        }
+        let (reply_tx, reply_rx) = oneshot::channel();
+        self.tx
+            .send(build(reply_tx))
+            .map_err(|_| WorkerError::Gone {
+                device_index: self.device_index,
+            })?;
+        match reply_rx.await {
+            Ok(result) => result.map_err(WorkerError::from),
+            Err(_) => Err(WorkerError::Gone {
+                device_index: self.device_index,
+            }),
+        }
+    }
+
+    /// TP mirror of [`Self::forward_logits_multi`] (#96): one CPU
+    /// `[vocab]` row per position from the leader's shard.
+    #[cfg(feature = "cuda")]
+    pub async fn tp_forward_logits_multi(
+        &self,
+        handle: TpHandle,
+        tokens: Vec<u32>,
+        offset: usize,
+    ) -> Result<Vec<Vec<f32>>, WorkerError> {
+        if self.poisoned.load(Ordering::Acquire) {
+            return Err(WorkerError::Poisoned {
+                device_index: self.device_index,
+            });
+        }
+        let (reply_tx, reply_rx) = oneshot::channel();
+        self.tx
+            .send(Job::TpForwardLogitsMulti {
+                handle,
+                tokens,
+                offset,
+                reply: reply_tx,
+            })
+            .map_err(|_| WorkerError::Gone {
+                device_index: self.device_index,
+            })?;
+        match reply_rx.await {
+            Ok(result) => result.map_err(WorkerError::from),
+            Err(_) => Err(WorkerError::Gone {
+                device_index: self.device_index,
+            }),
+        }
+    }
+
     #[cfg(feature = "cuda")]
     pub async fn tp_forward_logits(
         &self,

crates/neuron/src/harness/engine.rs

@@ -663,6 +663,7 @@ async fn finish_slot(slot: &mut Slot, reason: FinishReason, rate: &PrefillRateEm
                 prefill_ms: slot.prefill_ms,
                 decode_ms: slot.decode_start.elapsed().as_millis() as u32,
                 prefill_tokens: slot.prefill_tokens,
+                phases: None,
             },
         })
         .await;
@@ -1151,6 +1152,7 @@ mod tests {
                     prefill_ms: 6000,
                     decode_ms: 900_000,
                     prefill_tokens: 7413,
+                    phases: None,
                 },
             })
             .await

crates/neuron/src/harness/mod.rs

@@ -6,10 +6,13 @@ pub mod candle;
 pub mod capability;
 pub mod chat_template;
 pub mod context_limit;
+pub mod decision;
+pub mod decision_model;
 pub mod device_worker;
 pub mod engine;
 pub mod image;
 pub mod isq_cache;
+pub mod lang_detect;
 pub mod prefix_cache;
 pub mod preflight;
 pub mod preprocess;

crates/neuron/src/harness/preflight.rs

@@ -54,6 +54,11 @@ pub enum SourceFormat {
     /// Takes precedence over the safetensors sniff because diffusers
     /// repos also contain `.safetensors` shards (in subdirs).
     Diffusers,
+    /// A Laya decision-model checkpoint (#336): `rl_agent_config.json`
+    /// plus `encoder/config.json` at the root, no top-level
+    /// `config.json`. Served by the decision path — single device,
+    /// unquantized. Takes precedence over the safetensors sniff.
+    DecisionModel,
 }
 
 /// Output of `preflight` for a load that can proceed. Carries the
@@ -115,6 +120,12 @@ pub enum PreflightError {
         available: Vec<String>,
         nearest: Option<String>,
     },
+
+    /// The repo's format cannot be served with an option the spec
+    /// asked for (e.g. `tensor_parallel` or `quant` on a decision
+    /// model). `reason` says which and what to set instead.
+    #[error("cannot load '{model_id}' as requested: {reason}")]
+    Unsupported { model_id: String, reason: String },
 }
 
 /// Run the placement check.
@@ -230,6 +241,24 @@ pub async fn preflight(
         // The architecture compatibility check stays where it is —
         // `check_dense_config_supported` runs once `config.json` is
         // on disk, since it needs the parsed JSON.
+        // Decision models (#336) load unquantized on one device.
+        (SourceFormat::DecisionModel, tp, _) if tp > 1 => Err(PreflightError::Unsupported {
+            model_id: source_id.to_string(),
+            reason: format!(
+                "decision models load on a single device; tensor_parallel={tp} is not supported"
+            ),
+        }),
+        (SourceFormat::DecisionModel, _, Some(q)) => Err(PreflightError::Unsupported {
+            model_id: source_id.to_string(),
+            reason: format!("decision models load unquantized; remove quant={q:?}"),
+        }),
+        (SourceFormat::DecisionModel, _, None) => Ok(PlacementPlan {
+            model_id: source_id.to_string(),
+            format: format.clone(),
+            tp_size,
+            picked_quant_file: None,
+        }),
+
         (SourceFormat::DenseSafetensors { .. } | SourceFormat::Mixed { .. }, _, _) => {
             Ok(PlacementPlan {
                 model_id: source_id.to_string(),
@@ -295,6 +324,9 @@ pub fn classify(filenames: &[&str]) -> SourceFormat {
     if crate::harness::image::is_diffusers_layout(filenames) {
         return SourceFormat::Diffusers;
     }
+    if is_decision_layout(filenames) {
+        return SourceFormat::DecisionModel;
+    }
     let mut gguf_quants: Vec<String> = filenames
         .iter()
         .filter(|f| f.to_lowercase().ends_with(".gguf"))
@@ -318,6 +350,16 @@ pub fn classify(filenames: &[&str]) -> SourceFormat {
     }
 }
 
+/// A Laya decision checkpoint at the repo root: its decision config and
+/// its encoder config, with no top-level `config.json` (a causal LM
+/// that happened to ship a file with the same name still classifies as
+/// a text model).
+pub fn is_decision_layout(filenames: &[&str]) -> bool {
+    filenames.contains(&"rl_agent_config.json")
+        && filenames.contains(&"encoder/config.json")
+        && !filenames.contains(&"config.json")
+}
+
 /// Mirror of the quant-matching logic in `candle.rs::resolve_files` so
 /// preflight picks the same file the downstream loader would. Empty
 /// quant returns the first `.gguf` (any quant). Lowercased substring
@@ -651,6 +693,20 @@ mod tests {
                 tp_size,
                 picked_quant_file: None,
             }),
+            (SourceFormat::DecisionModel, tp, _) if tp > 1 => Err(PreflightError::Unsupported {
+                model_id: source_id.to_string(),
+                reason: format!("decision models load on a single device; tensor_parallel={tp}"),
+            }),
+            (SourceFormat::DecisionModel, _, Some(q)) => Err(PreflightError::Unsupported {
+                model_id: source_id.to_string(),
+                reason: format!("decision models load unquantized; remove quant={q:?}"),
+            }),
+            (SourceFormat::DecisionModel, _, None) => Ok(PlacementPlan {
+                model_id: source_id.to_string(),
+                format: format.clone(),
+                tp_size,
+                picked_quant_file: None,
+            }),
             (SourceFormat::DenseSafetensors { .. } | SourceFormat::Mixed { .. }, _, _) => {
                 Ok(PlacementPlan {
                     model_id: source_id.to_string(),
@@ -674,6 +730,66 @@ mod tests {
         assert_eq!(classify(&files), SourceFormat::Diffusers);
     }
 
+    /// The released `convaiinnovations/laya` listing: the English
+    /// checkpoint at the root, two more in subfolders, and no root
+    /// `config.json`.
+    const LAYA_FILES: [&str; 12] = [
+        "README.md",
+        "rl_agent_config.json",
+        "encoder/config.json",
+        "model.safetensors",
+        "tokenizer/tokenizer.json",
+        "tokenizer/tokenizer_config.json",
+        "multilingual/rl_agent_config.json",
+        "multilingual/encoder/config.json",
+        "multilingual/model.safetensors",
+        "typed-decisions/rl_agent_config.json",
+        "typed-decisions/encoder/config.json",
+        "typed-decisions/model.safetensors",
+    ];
+
+    #[test]
+    fn classify_laya_layout() {
+        assert_eq!(classify(&LAYA_FILES), SourceFormat::DecisionModel);
+        // A causal LM that also ships an rl config is still a text model.
+        let lm = [
+            "config.json",
+            "rl_agent_config.json",
+            "encoder/config.json",
+            "model.safetensors",
+        ];
+        assert!(matches!(
+            classify(&lm),
+            SourceFormat::DenseSafetensors { .. }
+        ));
+    }
+
+    #[test]
+    fn feasibility_decision_model() {
+        let fmt = classify(&LAYA_FILES);
+        let plan = decide(
+            &spec("convaiinnovations/laya", None, None),
+            &fmt,
+            &LAYA_FILES,
+        )
+        .expect("single-device unquantized decision load should pass");
+        assert_eq!(plan.format, SourceFormat::DecisionModel);
+        let err = decide(
+            &spec("convaiinnovations/laya", Some(2), None),
+            &fmt,
+            &LAYA_FILES,
+        )
+        .expect_err("tp>1 must be refused");
+        assert!(matches!(err, PreflightError::Unsupported { .. }));
+        let err = decide(
+            &spec("convaiinnovations/laya", None, Some("q8_0")),
+            &fmt,
+            &LAYA_FILES,
+        )
+        .expect_err("quant must be refused");
+        assert!(matches!(err, PreflightError::Unsupported { .. }));
+    }
+
     #[test]
     fn feasibility_diffusers_single_gpu_ok() {
         let files = ["model_index.json", "transformer/config.json"];

crates/neuron/src/harness/testdata/laya/lang.json

@@ -0,0 +1,347 @@
+{
+ "header": {
+  "laya_version": "0.3.21",
+  "torch_version": "2.14.0+cu130",
+  "device": "cpu",
+  "dtype": "float32",
+  "checkpoints": {
+   "english": {
+    "repo": "convaiinnovations/laya",
+    "revision": "55cf4c4ebb4ebe31b2550e8bdf3bd21b99753851",
+    "cfg": {
+     "max_len": 512,
+     "head_max_len": 192,
+     "temperature": [
+      1.6369030475616455,
+      1.2514300346374512,
+      1.983399510383606
+     ],
+     "temperature_by_options": {
+      "choice:3-5": 1.7601518630981445,
+      "choice:6-10": 1.0000158548355103,
+      "score:3-5": 1.2514300346374512,
+      "noul:2": 1.983399510383606,
+      "choice:11+": 0.10058280825614929,
+      "choice:2": 1.9063563346862793
+     }
+    },
+    "special_ids": {
+     "cls": 50281,
+     "sep": 50282,
+     "mask": 50284,
+     "pad": 50283
+    }
+   },
+   "multilingual": {
+    "repo": "convaiinnovations/laya/multilingual",
+    "revision": "55cf4c4ebb4ebe31b2550e8bdf3bd21b99753851",
+    "cfg": {
+     "max_len": 1024,
+     "head_max_len": 256,
+     "temperature": [
+      1.0,
+      1.0,
+      1.0
+     ],
+     "temperature_by_options": {}
+    },
+    "special_ids": {
+     "cls": 2,
+     "sep": 1,
+     "mask": 4,
+     "pad": 0
+    }
+   },
+   "typed-decisions": {
+    "repo": "convaiinnovations/laya/typed-decisions",
+    "revision": "55cf4c4ebb4ebe31b2550e8bdf3bd21b99753851",
+    "cfg": {
+     "max_len": 1024,
+     "head_max_len": 256,
+     "temperature": [
+      1.0148024559020996,
+      1.0374259948730469,
+      1.0575125217437744
+     ],
+     "temperature_by_options": {
+      "choice:3-5": 1.7601518630981445,
+      "choice:6-10": 1.0000158548355103,
+      "score:3-5": 1.2514300346374512,
+      "noul:2": 1.983399510383606,
+      "choice:11+": 0.10058280825614929,
+      "choice:2": 1.9063563346862793
+     }
+    },
+    "special_ids": {
+     "cls": 50281,
+     "sep": 50282,
+     "mask": 50284,
+     "pad": 50283
+    }
+   }
+  }
+ },
+ "cases": [
+  {
+   "state": "I was charged twice this month, I want my money back",
+   "analysis": {
+    "script": "latin",
+    "script_profile": {
+     "latin": 1.0
+    },
+    "language": "en",
+    "is_english": true,
+    "language_undecided": false,
+    "diacritic_rate": 0.0,
+    "non_latin_fraction": 0.0,
+    "mixed_segment": null
+   }
+  },
+  {
+   "state": "Me han cobrado dos veces este mes",
+   "analysis": {
+    "script": "latin",
+    "script_profile": {
+     "latin": 1.0
+    },
+    "language": "es",
+    "is_english": false,
+    "language_undecided": false,
+    "diacritic_rate": 0.0,
+    "non_latin_fraction": 0.0,
+    "mixed_segment": null
+   }
+  },
+  {
+   "state": "Die Rechnung ist falsch, bitte korrigieren",
+   "analysis": {
+    "script": "latin",
+    "script_profile": {
+     "latin": 1.0
+    },
+    "language": "de",
+    "is_english": false,
+    "language_undecided": false,
+    "diacritic_rate": 0.0,
+    "non_latin_fraction": 0.0,
+    "mixed_segment": null
+   }
+  },
+  {
+   "state": "Je voudrais annuler ma commande",
+   "analysis": {
+    "script": "latin",
+    "script_profile": {
+     "latin": 1.0
+    },
+    "language": "fr",
+    "is_english": false,
+    "language_undecided": false,
+    "diacritic_rate": 0.0,
+    "non_latin_fraction": 0.0,
+    "mixed_segment": null
+   }
+  },
+  {
+   "state": "मुझे पैसे वापस चाहिए",
+   "analysis": {
+    "script": "devanagari",
+    "script_profile": {
+     "devanagari": 1.0
+    },
+    "language": null,
+    "is_english": false,
+    "language_undecided": true,
+    "diacritic_rate": 0.0,
+    "non_latin_fraction": 1.0,
+    "mixed_segment": null
+   }
+  },
+  {
+   "state": "ខ្ញុំចង់បានប្រាក់វិញ",
+   "analysis": {
+    "script": "khmer",
+    "script_profile": {
+     "khmer": 1.0
+    },
+    "language": null,
+    "is_english": false,
+    "language_undecided": true,
+    "diacritic_rate": 0.0,
+    "non_latin_fraction": 1.0,
+    "mixed_segment": null
+   }
+  },
+  {
+   "state": "这个月我被扣了两次钱",
+   "analysis": {
+    "script": "han",
+    "script_profile": {
+     "han": 1.0
+    },
+    "language": null,
+    "is_english": false,
+    "language_undecided": true,
+    "diacritic_rate": 0.0,
+    "non_latin_fraction": 1.0,
+    "mixed_segment": null
+   }
+  },
+  {
+   "state": "注文をキャンセルしたいです",
+   "analysis": {
+    "script": "kana",
+    "script_profile": {
+     "han": 0.15384615384615385,
+     "kana": 0.8461538461538461
+    },
+    "language": null,
+    "is_english": false,
+    "language_undecided": true,
+    "diacritic_rate": 0.0,
+    "non_latin_fraction": 1.0,
+    "mixed_segment": null
+   }
+  },
+  {
+   "state": "تم خصم المبلغ مرتين",
+   "analysis": {
+    "script": "arabic",
+    "script_profile": {
+     "arabic": 1.0
+    },
+    "language": null,
+    "is_english": false,
+    "language_undecided": true,
+    "diacritic_rate": 0.0,
+    "non_latin_fraction": 1.0,
+    "mixed_segment": null
+   }
+  },
+  {
+   "state": "Мне дважды списали деньги",
+   "analysis": {
+    "script": "cyrillic",
+    "script_profile": {
+     "cyrillic": 1.0
+    },
+    "language": null,
+    "is_english": false,
+    "language_undecided": true,
+    "diacritic_rate": 0.0,
+    "non_latin_fraction": 1.0,
+    "mixed_segment": null
+   }
+  },
+  {
+   "state": "Order #5521 — 注文をキャンセルしたいです. Please help.",
+   "analysis": {
+    "script": "kana",
+    "script_profile": {
+     "latin": 0.5357142857142857,
+     "han": 0.07142857142857142,
+     "kana": 0.39285714285714285
+    },
+    "language": null,
+    "is_english": false,
+    "language_undecided": true,
+    "diacritic_rate": 0.0,
+    "non_latin_fraction": 0.4643,
+    "mixed_segment": null
+   }
+  },
+  {
+   "state": "ok",
+   "analysis": {
+    "script": "latin",
+    "script_profile": {
+     "latin": 1.0
+    },
+    "language": null,
+    "is_english": true,
+    "language_undecided": true,
+    "diacritic_rate": 0.0,
+    "non_latin_fraction": 0.0,
+    "mixed_segment": null
+   }
+  },
+  {
+   "state": "",
+   "analysis": {
+    "script": "unknown",
+    "script_profile": {},
+    "language": null,
+    "is_english": true,
+    "language_undecided": true,
+    "diacritic_rate": 0.0,
+    "non_latin_fraction": 0.0,
+    "mixed_segment": null
+   }
+  },
+  {
+   "state": "12345 67890",
+   "analysis": {
+    "script": "unknown",
+    "script_profile": {},
+    "language": null,
+    "is_english": true,
+    "language_undecided": true,
+    "diacritic_rate": 0.0,
+    "non_latin_fraction": 0.0,
+    "mixed_segment": null
+   }
+  },
+  {
+   "state": "café résumé naïve",
+   "analysis": {
+    "script": "latin",
+    "script_profile": {
+     "latin": 1.0
+    },
+    "language": null,
+    "is_english": false,
+    "language_undecided": true,
+    "diacritic_rate": 0.2353,
+    "non_latin_fraction": 0.0,
+    "mixed_segment": null
+   }
+  },
+  {
+   "state": {
+    "subject": "Invoice",
+    "body": "Bitte zahlen Sie bis Freitag"
+   },
+   "analysis": {
+    "script": "latin",
+    "script_profile": {
+     "latin": 1.0
+    },
+    "language": null,
+    "is_english": true,
+    "language_undecided": true,
+    "diacritic_rate": 0.0,
+    "non_latin_fraction": 0.0,
+    "mixed_segment": null
+   }
+  },
+  {
+   "state": [
+    "hello",
+    "bonjour",
+    "hola"
+   ],
+   "analysis": {
+    "script": "latin",
+    "script_profile": {
+     "latin": 1.0
+    },
+    "language": null,
+    "is_english": true,
+    "language_undecided": true,
+    "diacritic_rate": 0.0,
+    "non_latin_fraction": 0.0,
+    "mixed_segment": null
+   }
+  }
+ ]
+}

crates/neuron/src/harness/testdata/laya/tiny/encoder/config.json

@@ -0,0 +1,52 @@
+{
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "bos_token_id": 1,
+  "classifier_activation": "gelu",
+  "classifier_bias": false,
+  "classifier_dropout": 0.0,
+  "classifier_pooling": "cls",
+  "cls_token_id": 1,
+  "decoder_bias": true,
+  "deterministic_flash_attn": false,
+  "embedding_dropout": 0.0,
+  "eos_token_id": 2,
+  "global_attn_every_n_layers": 3,
+  "hidden_activation": "gelu",
+  "hidden_size": 32,
+  "initializer_cutoff_factor": 2.0,
+  "initializer_range": 0.3,
+  "intermediate_size": 48,
+  "layer_types": [
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention"
+  ],
+  "local_attention": 8,
+  "max_position_embeddings": 128,
+  "mlp_bias": false,
+  "mlp_dropout": 0.0,
+  "model_type": "modernbert",
+  "norm_bias": false,
+  "norm_eps": 1e-05,
+  "num_attention_heads": 2,
+  "num_hidden_layers": 4,
+  "pad_token_id": 0,
+  "rope_parameters": {
+    "full_attention": {
+      "rope_theta": 1000.0,
+      "rope_type": "default"
+    },
+    "sliding_attention": {
+      "rope_theta": 300.0,
+      "rope_type": "default"
+    }
+  },
+  "sep_token_id": 2,
+  "sparse_pred_ignore_index": -100,
+  "sparse_prediction": false,
+  "tie_word_embeddings": true,
+  "transformers_version": "5.17.0",
+  "vocab_size": 97
+}

crates/neuron/src/harness/testdata/laya/tiny/expected.json

@@ -0,0 +1,182 @@
+{
+ "seed": 7,
+ "seqs": [
+  {
+   "input_ids": [
+    1,
+    93,
+    82,
+    87,
+    51,
+    46,
+    89,
+    6,
+    68,
+    94,
+    86,
+    43,
+    23,
+    70,
+    42,
+    60,
+    35,
+    3,
+    86,
+    7,
+    20,
+    18,
+    90,
+    84,
+    62,
+    30,
+    42,
+    65,
+    34,
+    37,
+    30,
+    67,
+    2,
+    34,
+    77,
+    96,
+    24,
+    89,
+    19,
+    33
+   ],
+   "markers": [
+    3,
+    9,
+    17
+   ],
+   "qtype": 0,
+   "logits": [
+    -0.4152399003505707,
+    -0.5505563020706177,
+    -0.600245475769043
+   ],
+   "act": [
+    0.4727683663368225,
+    0.5272316336631775
+   ]
+  },
+  {
+   "input_ids": [
+    4,
+    5,
+    43,
+    95,
+    77,
+    35,
+    61,
+    13,
+    42,
+    24,
+    21,
+    7,
+    11,
+    38,
+    80,
+    54,
+    4,
+    29,
+    5,
+    80,
+    54,
+    78,
+    46
+   ],
+   "markers": [
+    4,
+    11
+   ],
+   "qtype": 1,
+   "logits": [
+    -0.1313231736421585,
+    -0.3265465199947357
+   ],
+   "act": [
+    0.40568628907203674,
+    0.5943136811256409
+   ]
+  },
+  {
+   "input_ids": [
+    67,
+    40,
+    83,
+    41,
+    77,
+    96,
+    58,
+    39,
+    76
+   ],
+   "markers": [
+    2
+   ],
+   "qtype": 2,
+   "logits": [
+    0.534652590751648
+   ],
+   "act": [
+    0.4226877987384796,
+    0.5773122310638428
+   ]
+  },
+  {
+   "input_ids": [
+    96,
+    32,
+    60,
+    44,
+    35,
+    47,
+    45,
+    51,
+    60,
+    51,
+    73,
+    50,
+    25,
+    79,
+    74,
+    24,
+    83,
+    7,
+    9,
+    91,
+    21,
+    87,
+    29,
+    25,
+    74,
+    18,
+    31,
+    46,
+    46,
+    30,
+    52
+   ],
+   "markers": [
+    5,
+    7,
+    12,
+    20,
+    26
+   ],
+   "qtype": 0,
+   "logits": [
+    -0.4001395106315613,
+    -0.5266168117523193,
+    -0.44161877036094666,
+    -0.43973028659820557,
+    -0.5682165622711182
+   ],
+   "act": [
+    0.46282079815864563,
+    0.537179172039032
+   ]
+  }
+ ]
+}

crates/neuron/src/harness/testdata/laya/tiny/rl_agent_config.json

@@ -0,0 +1,16 @@
+{
+ "encoder": "tiny-modernbert",
+ "head_layers": 2,
+ "max_len": 64,
+ "head_max_len": 24,
+ "act_costs": {
+  "escalate": 0.5
+ },
+ "amp_dtype": "bf16",
+ "temperature": [
+  1.0,
+  1.0,
+  1.0
+ ],
+ "temperature_by_options": {}
+}

crates/neuron/src/harness/testdata/laya/tiny/tokenizer.json

@@ -0,0 +1 @@
+{"version": "1.0", "truncation": null, "padding": null, "added_tokens": [], "normalizer": null, "pre_tokenizer": {"type": "Whitespace"}, "post_processor": null, "decoder": null, "model": {"type": "WordLevel", "vocab": {"[PAD]": 0, "[UNK]": 1, "[CLS]": 2, "[SEP]": 3, "[MASK]": 4, "a": 5, "b": 6}, "unk_token": "[UNK]"}}

crates/neuron/src/harness/testdata/laya/tiny/tokenizer_config.json

@@ -0,0 +1,7 @@
+{
+ "cls_token": "[CLS]",
+ "sep_token": "[SEP]",
+ "mask_token": "[MASK]",
+ "pad_token": "[PAD]",
+ "unk_token": "[UNK]"
+}

crates/neuron/src/harness/testdata/laya/tokenizer/tokenizer_config.json

@@ -0,0 +1,14 @@
+{
+  "clean_up_tokenization_spaces": true,
+  "cls_token": "[CLS]",
+  "mask_token": "[MASK]",
+  "model_input_names": [
+    "input_ids",
+    "attention_mask"
+  ],
+  "model_max_length": 8192,
+  "pad_token": "[PAD]",
+  "sep_token": "[SEP]",
+  "tokenizer_class": "PreTrainedTokenizerFast",
+  "unk_token": "[UNK]"
+}

crates/neuron/src/harness/tp/mod.rs

@@ -67,6 +67,22 @@ impl TpLeaderModel {
         }
     }
 
+    /// Logits at every position (#96). Only the `qwen3_5` arch has a
+    /// multi-position path; the older dense `qwen3` arch says so rather
+    /// than silently returning the last row.
+    pub fn forward_multi(
+        &mut self,
+        input: &candle_core::Tensor,
+        offset: usize,
+    ) -> candle_core::Result<candle_core::Tensor> {
+        match self {
+            TpLeaderModel::Qwen3_5(m) => m.forward_multi(input, offset),
+            TpLeaderModel::Qwen3(_) => Err(candle_core::Error::Msg(
+                "forward_multi: the tp qwen3 arch has no multi-position logits path".into(),
+            )),
+        }
+    }
+
     /// Chunked image prefill on rank 0. Only the vision-capable
     /// `qwen3_5` arch supports it; the dense `qwen3` arch has no tower.
     pub fn prefill_with_images_chunked(
@@ -996,6 +1012,112 @@ impl WorkerPool {
         }
     }
 
+    /// Multi-position forward across every rank (#96).
+    ///
+    /// Same fan-out / leader-forward / always-drain shape as
+    /// [`Self::generate_step`], and for the same reason: the
+    /// row-parallel `AllReduce`s only complete when every rank has
+    /// issued them, and a leader error must still drain the workers or
+    /// the stale replies poison the next request.
+    ///
+    /// Returns one `[vocab]` row per position from the leader's rank-0
+    /// shard, which is the whole answer because `lm_head` is
+    /// replicated.
+    #[cfg(feature = "cuda")]
+    pub async fn generate_step_multi(
+        &mut self,
+        model_id: &str,
+        leader_handle: super::device_worker::TpHandle,
+        tokens: Vec<u32>,
+        offset: usize,
+    ) -> Result<Vec<Vec<f32>>> {
+        let step_start = std::time::Instant::now();
+        let tokens_len = tokens.len();
+        tracing::debug!(
+            model = %model_id,
+            tokens = tokens_len,
+            offset,
+            "WorkerPool::generate_step_multi: fan-out"
+        );
+        for w in &mut self.workers {
+            w.send_only(&WorkerRequest::GenerateStepMulti {
+                model_id: model_id.to_string(),
+                tokens: tokens.clone(),
+                offset,
+            })
+            .await?;
+        }
+
+        let leader_start = std::time::Instant::now();
+        let timeout = tp_step_timeout();
+        let leader_fut = self
+            .leader_worker
+            .tp_forward_logits_multi(leader_handle, tokens, offset);
+        let leader_result = match tokio::time::timeout(timeout, leader_fut).await {
+            Ok(r) => r,
+            Err(_elapsed) => {
+                self.watchdog_abort_leader_comm(model_id, timeout.as_secs());
+                anyhow::bail!(
+                    "tp watchdog: leader multi-position forward exceeded {}s deadline; \
+                     aborted wedged NCCL comm — model will auto-recover",
+                    timeout.as_secs()
+                );
+            }
+        };
+        let leader_ok = leader_result.is_ok();
+        if !leader_ok {
+            let detail = leader_result
+                .as_ref()
+                .err()
+                .map(|e| format!("{e:#}"))
+                .unwrap_or_default();
+            tracing::warn!(
+                model = %model_id,
+                tokens = tokens_len,
+                offset,
+                leader_ms = leader_start.elapsed().as_millis(),
+                error = %detail,
+                "WorkerPool::generate_step_multi: leader forward failed"
+            );
+        }
+
+        let worker_errors = drain_workers(&mut self.workers, |r| match r {
+            WorkerResponse::GenerateStepOk => Ok(()),
+            WorkerResponse::Error { kind, message } => Err(format!("[{kind}]: {message}")),
+            other => Err(format!("expected GenerateStepOk, got {other:?}")),
+        })
+        .await;
+        tracing::debug!(
+            model = %model_id,
+            errors = worker_errors.len(),
+            total_ms = step_start.elapsed().as_millis(),
+            "WorkerPool::generate_step_multi: workers drained"
+        );
+
+        match leader_result {
+            Ok(values) => {
+                if worker_errors.is_empty() {
+                    Ok(values)
+                } else {
+                    anyhow::bail!(
+                        "GenerateStepMulti: leader succeeded but workers failed: {}",
+                        worker_errors.join("; ")
+                    )
+                }
+            }
+            Err(e) => {
+                if worker_errors.is_empty() {
+                    Err(anyhow::Error::new(e).context("GenerateStepMulti: leader forward failed"))
+                } else {
+                    Err(anyhow::Error::new(e).context(format!(
+                        "GenerateStepMulti: leader forward failed and workers also failed: {}",
+                        worker_errors.join("; ")
+                    )))
+                }
+            }
+        }
+    }
+
     /// One lockstep batched decode step across every rank (#98).
     /// Same fan-out / leader-forward / always-drain shape as
     /// [`Self::generate_step`]; every rank derives positions + mask

crates/neuron/src/harness/tp/rpc.rs

@@ -78,6 +78,18 @@ pub enum WorkerRequest {
     /// same op. The leader does *not* receive logits back over RPC; it
     /// runs its own rank-0 forward in parallel and uses its own logits
     /// for sampling.
+    /// Multi-position forward (#96): like `GenerateStep`, but every
+    /// rank keeps the logits at *all* `tokens.len()` positions rather
+    /// than the last. The collectives are identical, so the ranks stay
+    /// in lockstep; only the leader's logits are read. Worker replies
+    /// with `GenerateStepOk`.
+    GenerateStepMulti {
+        model_id: String,
+        /// Identical on every rank, as for `GenerateStep`.
+        tokens: Vec<u32>,
+        offset: usize,
+    },
+
     GenerateStep {
         model_id: String,
         /// Input token ids for this step. For prefill, the whole prompt;

crates/neuron/src/harness/tp/tp_qwen3_5.rs

@@ -1881,6 +1881,27 @@ impl TpQwen3_5Model {
     }
 }
 
+/// Additive causal mask for the draft head, `(1, 1, l, l + offset)`.
+/// The head keeps its own KV, so its mask is built from its own
+/// position count rather than the target's.
+fn mtp_causal_mask(
+    l: usize,
+    offset: usize,
+    dtype: DType,
+    dev: &Device,
+) -> candle_core::Result<Tensor> {
+    let total = l + offset;
+    let mut data = vec![0f32; l * total];
+    for i in 0..l {
+        for j in 0..total {
+            if j > i + offset {
+                data[i * total + j] = f32::NEG_INFINITY;
+            }
+        }
+    }
+    Tensor::from_vec(data, (1, 1, l, total), dev)?.to_dtype(dtype)
+}
+
 pub struct TpQwen3_5ForCausalLM {
     base: TpQwen3_5Model,
     lm_head: super::tp_linear::MaybeQuantLinear,
@@ -1894,6 +1915,23 @@ pub struct TpQwen3_5ForCausalLM {
     /// `<|image_pad|>` sentinel id (mirrors `Config::image_token_id`);
     /// the splice target for `forward_with_vision`.
     image_token_id: Option<u32>,
+    /// The MTP draft head (#96), on the **leader only**.
+    ///
+    /// Unsharded, unlike every other weight here, and that is correct
+    /// rather than lazy: after each row-parallel `AllReduce` every rank
+    /// holds the *same* hidden state, and `lm_head` is replicated. So a
+    /// head fed the leader's hidden produces exactly what a sharded one
+    /// would, with no collectives — which also means drafting cannot
+    /// wedge the ranks, since the other ranks are not in it.
+    ///
+    /// `None` on the subprocess ranks and on any load that did not ask
+    /// for it. Loaded by [`Self::load_mtp_head`], never by `load`.
+    mtp: Option<crate::harness::arch::qwen3_5::mtp::MtpHead>,
+    /// The last forward's hidden state, `(1, L, hidden)` — what the
+    /// draft head consumes. Kept because the head runs *after* the
+    /// target's step, outside the collective, so it cannot recompute
+    /// it. One `Arc` clone per forward; no copy.
+    last_hidden: Option<Tensor>,
 }
 
 /// Load the replicated vision tower from the unsharded `model.visual.*`
@@ -1944,6 +1982,8 @@ impl TpQwen3_5ForCausalLM {
             lm_head,
             vision,
             image_token_id,
+            mtp: None,
+            last_hidden: None,
         };
         log_construction_complete(cfg, rank, world_size, quant, model.device());
         Ok(model)
@@ -1969,6 +2009,8 @@ impl TpQwen3_5ForCausalLM {
             lm_head,
             vision,
             image_token_id,
+            mtp: None,
+            last_hidden: None,
         };
         log_construction_complete(cfg, rank, world_size, quant, model.device());
         Ok(model)
@@ -2003,9 +2045,201 @@ impl TpQwen3_5ForCausalLM {
     pub fn forward(&mut self, input: &Tensor, offset: usize) -> candle_core::Result<Tensor> {
         let (_, l) = input.dims2()?;
         let hidden = self.base.forward(input, offset)?;
+        self.last_hidden = Some(hidden.clone());
         hidden.i((.., l - 1.., ..))?.apply(&self.lm_head)
     }
 
+    /// Logits at **every** position: `(B, L, vocab)`.
+    ///
+    /// The TP mirror of `Qwen3_5ForCausalLM::forward_multi` (#96). A
+    /// speculative verify pass forwards the drafted block in one pass
+    /// and needs the target's own token at each drafted position;
+    /// [`Self::forward`] slices the last one off before the LM head.
+    ///
+    /// Every rank runs this, because the row-parallel `o_proj` and the
+    /// MLP's `down_proj` issue `AllReduce` collectives that only
+    /// complete when all ranks arrive. The leader's copy is the one
+    /// whose logits are read — `lm_head` is replicated, so rank 0's
+    /// answer is the whole answer.
+    /// Load the MTP draft head onto this rank, unsharded (#96).
+    ///
+    /// Only the leader should call this. See the `mtp` field for why an
+    /// unsharded head is numerically right; the practical consequence
+    /// is that drafting issues no collectives and therefore cannot
+    /// leave the other ranks waiting.
+    pub fn load_mtp_head(&mut self, cfg: &TextConfig, vb: &ShardedVarBuilder) -> Result<()> {
+        let head = crate::harness::arch::qwen3_5::mtp::MtpHead::load(
+            cfg,
+            Arc::clone(&self.base.rotary),
+            vb,
+        )
+        .context("load mtp head on the tp leader")?;
+        self.mtp = Some(head);
+        Ok(())
+    }
+
+    /// Whether this rank carries a draft head.
+    pub fn has_mtp(&self) -> bool {
+        self.mtp.is_some()
+    }
+
+    /// Walk the draft head over a prompt chunk, using the hidden state
+    /// the target's forward just produced.
+    ///
+    /// The head trails the target by one token: at position `i` it
+    /// consumes the embedding of token `i+1` with the target's hidden
+    /// state at `i`. `shifted` is therefore the chunk's tokens moved
+    /// one left, and must be one shorter than the retained hidden block
+    /// unless the caller has the next chunk's first token.
+    pub fn mtp_prefill_chunk(&mut self, shifted: &[u32], start_pos: usize) -> Result<()> {
+        let hidden = self
+            .last_hidden
+            .as_ref()
+            .context("mtp_prefill_chunk: no hidden state retained; run the target forward first")?
+            .clone();
+        let l = shifted.len();
+        anyhow::ensure!(l > 0, "mtp_prefill_chunk: empty chunk");
+        let (_, have, _) = hidden.dims3()?;
+        anyhow::ensure!(
+            have >= l,
+            "mtp_prefill_chunk: {l} tokens against {have} retained hidden positions"
+        );
+        let device = self.base.device.clone();
+        let dtype = self.base.dtype;
+        let rotary = Arc::clone(&self.base.rotary);
+        let tokens = Tensor::new(shifted, &device)?.unsqueeze(0)?;
+        let embeds = self.base.embed_tokens.forward(&tokens)?;
+        let hid = hidden.i((.., ..l, ..))?;
+        let mask = mtp_causal_mask(l, start_pos, dtype, &device)?;
+        let positions: Vec<usize> = (start_pos..start_pos + l).collect();
+        let (cos, sin) = rotary.cos_sin_at(&positions)?;
+        let head = self
+            .mtp
+            .as_mut()
+            .context("mtp_prefill_chunk: no draft head on this rank")?;
+        head.forward(&embeds, &hid, Some(&mask), &cos, &sin)?;
+        Ok(())
+    }
+
+    /// Draft `k` tokens, then put the head's cache back.
+    ///
+    /// The head is walked over tokens that may be rejected, so the
+    /// snapshot/restore pair around the chain is what keeps this an
+    /// observation rather than a commitment. Returns the drafted token
+    /// ids, argmax-sampled through the target's own (replicated)
+    /// `lm_head`.
+    pub fn mtp_draft(&mut self, first_token: u32, start_pos: usize, k: usize) -> Result<Vec<u32>> {
+        anyhow::ensure!(k > 0, "mtp_draft: k must be positive");
+        let mut hidden = self
+            .last_hidden
+            .as_ref()
+            .context("mtp_draft: no hidden state retained; run the target forward first")?
+            .clone();
+        let (_, have, _) = hidden.dims3()?;
+        if have > 1 {
+            hidden = hidden.i((.., have - 1.., ..))?;
+        }
+        let device = self.base.device.clone();
+        let rotary = Arc::clone(&self.base.rotary);
+
+        let snap = {
+            let head = self
+                .mtp
+                .as_ref()
+                .context("mtp_draft: no draft head on this rank")?;
+            head.snapshot_kv()?
+        };
+
+        let mut drafted = Vec::with_capacity(k);
+        let mut token = first_token;
+        let mut cur = hidden;
+        let mut err: Option<anyhow::Error> = None;
+        for step in 0..k {
+            let ids = match Tensor::new(&[token], &device).and_then(|t| t.unsqueeze(0)) {
+                Ok(t) => t,
+                Err(e) => {
+                    err = Some(e.into());
+                    break;
+                }
+            };
+            let out = (|| -> Result<Tensor> {
+                let embeds = self.base.embed_tokens.forward(&ids)?;
+                let (cos, sin) = rotary.cos_sin_at(&[start_pos + step])?;
+                let head = self
+                    .mtp
+                    .as_mut()
+                    .context("mtp_draft: no draft head on this rank")?;
+                head.forward(&embeds, &cur, None, &cos, &sin)
+            })();
+            let out = match out {
+                Ok(t) => t,
+                Err(e) => {
+                    err = Some(e);
+                    break;
+                }
+            };
+            let next = match self.lm_head.forward(&out).and_then(|logits| {
+                logits
+                    .i((0, 0, ..))?
+                    .to_dtype(DType::F32)?
+                    .argmax(candle_core::D::Minus1)?
+                    .to_scalar::<u32>()
+            }) {
+                Ok(t) => t,
+                Err(e) => {
+                    err = Some(e.into());
+                    break;
+                }
+            };
+            drafted.push(next);
+            token = next;
+            cur = out;
+        }
+
+        // Restore even when a step failed: a half-walked head is worse
+        // than no draft, because every later round inherits it.
+        if let Some(head) = self.mtp.as_mut() {
+            head.restore_kv(&snap)?;
+        }
+        match err {
+            Some(e) => Err(e),
+            None => Ok(drafted),
+        }
+    }
+
+    /// Walk the head over one token the target actually committed, so
+    /// its cache stays level with the target's.
+    pub fn mtp_advance(&mut self, token: u32, pos: usize) -> Result<()> {
+        let hidden = self
+            .last_hidden
+            .as_ref()
+            .context("mtp_advance: no hidden state retained")?
+            .clone();
+        let (_, have, _) = hidden.dims3()?;
+        let hidden = if have > 1 {
+            hidden.i((.., have - 1.., ..))?
+        } else {
+            hidden
+        };
+        let device = self.base.device.clone();
+        let rotary = Arc::clone(&self.base.rotary);
+        let ids = Tensor::new(&[token], &device)?.unsqueeze(0)?;
+        let embeds = self.base.embed_tokens.forward(&ids)?;
+        let (cos, sin) = rotary.cos_sin_at(&[pos])?;
+        let head = self
+            .mtp
+            .as_mut()
+            .context("mtp_advance: no draft head on this rank")?;
+        head.forward(&embeds, &hidden, None, &cos, &sin)?;
+        Ok(())
+    }
+
+    pub fn forward_multi(&mut self, input: &Tensor, offset: usize) -> candle_core::Result<Tensor> {
+        let hidden = self.base.forward(input, offset)?;
+        self.last_hidden = Some(hidden.clone());
+        hidden.apply(&self.lm_head)
+    }
+
     /// Lockstep batched decode step (#98): `(B, 1)` input, per-row
     /// positions, padding mask from
     /// [`TpQwen3_5Model::batch_decode_mask`]. Returns `(B, 1, vocab)`.

crates/neuron/src/harness/tp/worker.rs

@@ -52,6 +52,21 @@ impl WorkerModel {
         }
     }
 
+    /// Logits at every position (#96). This rank discards them; what
+    /// matters is the identical collective sequence.
+    fn forward_multi(
+        &mut self,
+        input: &candle_core::Tensor,
+        offset: usize,
+    ) -> candle_core::Result<candle_core::Tensor> {
+        match self {
+            WorkerModel::Qwen3_5(m) => m.forward_multi(input, offset),
+            WorkerModel::Qwen3(_) => Err(candle_core::Error::Msg(
+                "forward_multi: the tp qwen3 arch has no multi-position logits path".into(),
+            )),
+        }
+    }
+
     /// Chunked image prefill on this rank. Only the vision-capable
     /// `qwen3_5` arch has a replicated tower; the dense `qwen3` arch
     /// errors. The returned logits are discarded by the caller (the
@@ -269,6 +284,11 @@ impl WorkerState {
                 tokens,
                 offset,
             } => self.handle_generate_step(&model_id, tokens, offset),
+            WorkerRequest::GenerateStepMulti {
+                model_id,
+                tokens,
+                offset,
+            } => self.handle_generate_step_multi(&model_id, tokens, offset),
             WorkerRequest::GenerateStepWithImages {
                 model_id,
                 tokens,
@@ -594,6 +614,80 @@ impl WorkerState {
         }
     }
 
+    /// Multi-position forward (#96). Identical collectives to
+    /// `handle_generate_step` — the only difference is that the model
+    /// keeps every position's logits instead of the last. This rank
+    /// discards them either way; what matters is that it issues the
+    /// same `AllReduce` sequence, or the leader's forward never
+    /// completes.
+    #[cfg(feature = "cuda")]
+    fn handle_generate_step_multi(
+        &mut self,
+        model_id: &str,
+        tokens: Vec<u32>,
+        offset: usize,
+    ) -> WorkerResponse {
+        use candle_core::Tensor;
+
+        let Some(model) = self.models.get_mut(model_id) else {
+            return WorkerResponse::Error {
+                kind: "model_not_loaded".into(),
+                message: format!("model '{model_id}' not loaded on rank {}", self.config.rank),
+            };
+        };
+        let device = model.device().clone();
+        let input = match Tensor::new(tokens.as_slice(), &device).and_then(|t| t.unsqueeze(0)) {
+            Ok(t) => t,
+            Err(e) => {
+                return WorkerResponse::Error {
+                    kind: "forward_failed".into(),
+                    message: format!("build input tensor: {e}"),
+                };
+            }
+        };
+        let start = std::time::Instant::now();
+        tracing::debug!(
+            rank = self.config.rank,
+            model = %model_id,
+            tokens = tokens.len(),
+            offset,
+            "worker GenerateStepMulti: forward starting"
+        );
+        if let Err(e) = model.forward_multi(&input, offset) {
+            tracing::warn!(
+                rank = self.config.rank,
+                model = %model_id,
+                elapsed_ms = start.elapsed().as_millis(),
+                error = %e,
+                "worker GenerateStepMulti: forward failed"
+            );
+            return WorkerResponse::Error {
+                kind: "forward_failed".into(),
+                message: format!("TP multi-position forward: {e}"),
+            };
+        }
+        tracing::debug!(
+            rank = self.config.rank,
+            model = %model_id,
+            elapsed_ms = start.elapsed().as_millis(),
+            "worker GenerateStepMulti: forward done"
+        );
+        WorkerResponse::GenerateStepOk
+    }
+
+    #[cfg(not(feature = "cuda"))]
+    fn handle_generate_step_multi(
+        &mut self,
+        _model_id: &str,
+        _tokens: Vec<u32>,
+        _offset: usize,
+    ) -> WorkerResponse {
+        WorkerResponse::Error {
+            kind: "cuda_feature_not_enabled".into(),
+            message: "GenerateStepMulti requires --features cuda".into(),
+        }
+    }
+
     /// One lockstep batched decode step on this rank (#98). Mirrors
     /// the leader's `forward_logits_batch`: derives per-row positions
     /// and the padding mask locally from the broadcast geometry, runs

crates/neuron/src/lib.rs

@@ -5,6 +5,7 @@ pub mod cuda;
 pub mod discovery;
 pub mod harness;
 pub mod health;
+pub mod metrics;
 pub mod startup;
 pub mod version;
 pub mod wire;

crates/neuron/src/main.rs

@@ -220,6 +220,12 @@ async fn daemon(args: Args) -> Result<()> {
     // host look down to anything probing `/health` during pre-warm.
     // The pre-warm task runs in the background instead — `/health`
     // surfaces its progress via the activation field.
+    // Before the listener binds, so the very first scrape after a
+    // restart finds a recorder rather than a 503.
+    if let Err(e) = neuron::metrics::install() {
+        tracing::warn!(error = %e, "metrics recorder not installed; /metrics will 503");
+    }
+
     let app = api::neuron_routes().with_state(Arc::clone(&state));
     let addr: std::net::SocketAddr = format!("0.0.0.0:{port}").parse()?;
     let listener = tokio::net::TcpListener::bind(addr).await?;

crates/neuron/src/metrics.rs

@@ -0,0 +1,173 @@
+//! Prometheus metrics for neuron.
+//!
+//! neuron had no `/metrics` surface at all: every signal it produced
+//! reached Prometheus only by cortex polling `/health` every ~10 s and
+//! republishing it as gauges. That is enough for routing decisions and
+//! useless for asking where a request's time went — an EMA sampled at
+//! 10 s cannot see inside a one-second forward, and the finest split
+//! that existed anywhere was prefill versus decode.
+//!
+//! Served from the main API port rather than a second listener: neuron
+//! already runs an axum server, and a separate port is one more thing
+//! to open in the firewall for no benefit.
+
+use anyhow::Result;
+use metrics_exporter_prometheus::{Matcher, PrometheusBuilder, PrometheusHandle};
+use std::sync::OnceLock;
+
+static HANDLE: OnceLock<PrometheusHandle> = OnceLock::new();
+
+/// Install the process-wide recorder. Idempotent in effect: a second
+/// call leaves the first handle in place.
+pub fn install() -> Result<()> {
+    let handle = with_buckets(PrometheusBuilder::new())?
+        .install_recorder()
+        .map_err(|e| anyhow::anyhow!("failed to install Prometheus recorder: {e}"))?;
+    describe();
+    let _ = HANDLE.set(handle);
+    Ok(())
+}
+
+/// Render the current metrics, or `None` if the recorder was never
+/// installed. A `None` here must surface as an error rather than an
+/// empty body: an empty scrape is indistinguishable from a healthy
+/// process with nothing to report.
+pub fn render() -> Option<String> {
+    HANDLE.get().map(|h| h.render())
+}
+
+/// Explicit buckets, or `metrics-exporter-prometheus` renders every
+/// histogram as a **summary** — `{quantile=...}` series, no `_bucket`,
+/// no aggregation across hosts, and every quantile decaying to `0`
+/// when idle so "no traffic" looks like "instant". The same trap
+/// cortex's exporter documents; the fix has to be repeated per metric.
+fn with_buckets(builder: PrometheusBuilder) -> Result<PrometheusBuilder> {
+    // Per-request phase sums: microseconds for an emit on a fast model,
+    // minutes for the forward on a long turn.
+    let phase_seconds = &[
+        0.001, 0.005, 0.025, 0.1, 0.25, 1.0, 2.5, 5.0, 15.0, 30.0, 60.0, 300.0,
+    ];
+    let seconds_short = &[
+        0.05, 0.1, 0.25, 0.5, 1.0, 2.0, 5.0, 10.0, 20.0, 30.0, 60.0, 120.0,
+    ];
+    let seconds_long = &[
+        0.1, 0.5, 1.0, 2.5, 5.0, 10.0, 30.0, 60.0, 120.0, 300.0, 600.0, 1200.0,
+    ];
+    builder
+        .set_buckets_for_metric(
+            Matcher::Full("neuron_decode_phase_seconds".into()),
+            phase_seconds,
+        )?
+        .set_buckets_for_metric(
+            Matcher::Full("neuron_prefill_seconds".into()),
+            seconds_short,
+        )?
+        .set_buckets_for_metric(Matcher::Full("neuron_decode_seconds".into()), seconds_long)
+        .map_err(|e| anyhow::anyhow!("failed to configure histogram buckets: {e}"))
+}
+
+fn describe() {
+    metrics::describe_histogram!(
+        "neuron_decode_phase_seconds",
+        "Decode-loop time by phase, summed per request. The `_sum` \
+         series is the point: ratio one phase against the total to see \
+         where decode goes, without needing per-step resolution."
+    );
+    metrics::describe_histogram!("neuron_prefill_seconds", "Prefill wall-clock per request");
+    metrics::describe_histogram!("neuron_decode_seconds", "Decode wall-clock per request");
+    metrics::describe_counter!(
+        "neuron_completion_tokens_total",
+        "Tokens generated, the denominator for per-token phase cost"
+    );
+}
+
+/// Record one finished inference.
+///
+/// `phases` is optional because not every serving path measures them,
+/// and a path that does not must emit **nothing** rather than zeros —
+/// a zero sample pulls the histogram's mean down and reads as "this
+/// phase is free here", which is the opposite of "nobody looked".
+///
+/// The residual is emitted as its own phase. Whatever the brackets
+/// failed to cover is then visible as a labelled series instead of
+/// being silently divided among the phases that were measured.
+pub fn record_finish(
+    model: &str,
+    prefill_ms: u32,
+    decode_ms: u32,
+    completion_tokens: u32,
+    phases: Option<(u32, u32, u32)>,
+) {
+    let m = model.to_string();
+    metrics::histogram!("neuron_prefill_seconds", "model" => m.clone())
+        .record(prefill_ms as f64 / 1e3);
+    metrics::histogram!("neuron_decode_seconds", "model" => m.clone())
+        .record(decode_ms as f64 / 1e3);
+    metrics::counter!("neuron_completion_tokens_total", "model" => m.clone())
+        .increment(completion_tokens as u64);
+
+    let Some((forward_ms, sample_ms, emit_ms)) = phases else {
+        return;
+    };
+    for (phase, ms) in [
+        ("forward", forward_ms),
+        ("sample", sample_ms),
+        ("emit", emit_ms),
+    ] {
+        metrics::histogram!("neuron_decode_phase_seconds", "model" => m.clone(), "phase" => phase)
+            .record(ms as f64 / 1e3);
+    }
+    // Saturating: the phases are measured inside the decode window, so
+    // they cannot legitimately exceed it, but clock granularity can
+    // round them a millisecond over and an underflow here would report
+    // an enormous residual.
+    let residual =
+        decode_ms.saturating_sub(forward_ms.saturating_add(sample_ms).saturating_add(emit_ms));
+    metrics::histogram!("neuron_decode_phase_seconds", "model" => m, "phase" => "residual")
+        .record(residual as f64 / 1e3);
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    /// An unmeasured path must contribute no phase samples at all.
+    /// Zeros would drag the histogram's mean toward zero and read as
+    /// "this phase costs nothing on this model".
+    #[test]
+    fn phases_absent_means_no_phase_series() {
+        let handle = with_buckets(PrometheusBuilder::new())
+            .expect("buckets")
+            .install_recorder()
+            .expect("recorder");
+        describe();
+
+        record_finish("measured", 100, 1000, 64, Some((900, 50, 10)));
+        record_finish("unmeasured", 100, 1000, 64, None);
+        let out = handle.render();
+
+        assert!(
+            out.contains("neuron_decode_phase_seconds") && out.contains(r#"model="measured""#),
+            "measured model must emit phases:\n{out}"
+        );
+        for line in out
+            .lines()
+            .filter(|l| l.contains("neuron_decode_phase_seconds"))
+        {
+            assert!(
+                !line.contains(r#"model="unmeasured""#),
+                "unmeasured model must emit no phase series, got: {line}"
+            );
+        }
+        // Buckets, not a summary: `histogram_quantile()` needs `_bucket`.
+        assert!(
+            out.contains("neuron_decode_phase_seconds_bucket"),
+            "must export as a histogram, not a summary:\n{out}"
+        );
+        // 1000 - (900 + 50 + 10) = 40 ms of residual, as its own series.
+        assert!(
+            out.contains(r#"phase="residual""#),
+            "residual must be its own series:\n{out}"
+        );
+    }
+}

crates/neuron/src/startup.rs

@@ -158,6 +158,7 @@ fn preflight_kind(err: &PreflightError) -> &'static str {
         PreflightError::EmptyRepo { .. } => "empty_repo",
         PreflightError::TpRequiresSafetensors { .. } => "tp_requires_safetensors",
         PreflightError::QuantNotFound { .. } => "quant_not_found",
+        PreflightError::Unsupported { .. } => "unsupported_load_option",
     }
 }
 

crates/neuron/src/wire/event.rs

@@ -121,6 +121,41 @@ pub struct FinishTiming {
     /// the elapsed `prefill_ms` drops while this stays the full prompt
     /// length, so a high implied rate is itself the cache-hit signal.
     pub prefill_tokens: u32,
+    /// How the decode loop's wall-clock divides, when the serving path
+    /// measured it. `None` means "this path does not report phases",
+    /// which is deliberately distinct from "measured, and they were
+    /// zero" — a phase split that cannot express its own absence
+    /// invites reading an unmeasured path as an idle one.
+    pub phases: Option<PhaseTiming>,
+}
+
+/// Where a decode loop's wall-clock actually went, summed over every
+/// step of the request.
+///
+/// These are sums across steps, not a breakdown of one step, and each
+/// boundary is a point the loop already synchronises at: the forward
+/// hands back CPU-side logits, so no device synchronisation is added
+/// and the numbers are elapsed work rather than kernel-launch time.
+///
+/// The three should roughly account for `decode_ms`; a large residual
+/// is itself a finding (it is time inside the loop that none of the
+/// brackets cover).
+#[derive(Debug, Clone, Copy, Default, PartialEq, Eq)]
+pub struct PhaseTiming {
+    /// Submitting the step to the device worker and getting logits
+    /// back: the channel hop, the model forward, and the device-to-host
+    /// copy of the logits row. For a large vocabulary that copy is not
+    /// negligible -- 248,320 f32 is ~993 KB per step.
+    pub forward_ms: u32,
+    /// Wrapping the logits in a CPU tensor and sampling: repeat/presence
+    /// penalties plus the `LogitsProcessor`. Scales with vocabulary,
+    /// not with model size, and top-p sorts the distribution.
+    pub sample_ms: u32,
+    /// Detokenising the sampled token and handing the delta to the
+    /// consumer. **Includes backpressure**: the send awaits a bounded
+    /// channel, so a slow reader is charged here. This is emit cost,
+    /// not detokenisation cost, and the two are not separated.
+    pub emit_ms: u32,
 }
 
 /// Why a stream stopped. Stays small on purpose — anything that

crates/neuron/src/wire/mod.rs

@@ -22,6 +22,6 @@ pub mod openai_chat;
 pub mod openai_responses;
 
 pub use event::{
-    FinishReason, FinishTiming, InferenceEvent, ReasoningTokenPair, ToolCallTokenPair,
+    FinishReason, FinishTiming, InferenceEvent, PhaseTiming, ReasoningTokenPair, ToolCallTokenPair,
     detect_reasoning_token_pair, detect_tool_call_token_pair,
 };

crates/neuron/src/wire/openai_chat.rs

@@ -444,6 +444,9 @@ fn usage_chunk(
                 prefill_ms: t.prefill_ms as u64,
                 decode_ms: t.decode_ms as u64,
                 prefill_tokens: t.prefill_tokens as u64,
+                forward_ms: t.phases.map(|p| p.forward_ms as u64),
+                sample_ms: t.phases.map(|p| p.sample_ms as u64),
+                emit_ms: t.phases.map(|p| p.emit_ms as u64),
             }),
         }),
         extra: serde_json::Value::Object(Default::default()),
@@ -596,6 +599,9 @@ pub async fn collect_chat_completion(
                 prefill_ms: t.prefill_ms as u64,
                 decode_ms: t.decode_ms as u64,
                 prefill_tokens: t.prefill_tokens as u64,
+                forward_ms: t.phases.map(|p| p.forward_ms as u64),
+                sample_ms: t.phases.map(|p| p.sample_ms as u64),
+                emit_ms: t.phases.map(|p| p.emit_ms as u64),
             }),
         }),
         extra: serde_json::Value::Object(Default::default()),
@@ -685,6 +691,7 @@ mod tests {
         );
     }
     use super::*;
+    use crate::wire::event::PhaseTiming;
 
     // ── Non-streaming collector (#285) ────────────────────────────
 
@@ -799,6 +806,7 @@ mod tests {
                 prefill_ms: 12,
                 decode_ms: 34,
                 prefill_tokens: 100,
+                phases: None,
             }),
         })
         .await
@@ -813,6 +821,65 @@ mod tests {
         assert_eq!((t.prefill_ms, t.decode_ms, t.prefill_tokens), (12, 34, 100));
     }
 
+    /// The phase split must survive to the wire, and an unmeasured path
+    /// must serialise as **absent** rather than as zeros.
+    ///
+    /// That distinction is the reason the field is an `Option` at all:
+    /// a consumer reading `forward_ms: 0` off a path that never
+    /// measured would conclude the forward is free. The serialised
+    /// form is asserted, not just the struct, because
+    /// `skip_serializing_if` is what actually enforces it and a plain
+    /// field comparison cannot see it.
+    #[tokio::test]
+    async fn phase_timing_reaches_the_wire_and_absence_is_not_zero() {
+        async fn timing_json(phases: Option<PhaseTiming>) -> serde_json::Value {
+            let (tx, rx) = mpsc::channel(8);
+            tx.send(InferenceEvent::Finish {
+                reason: FinishReason::Stop,
+                prompt_tokens: 10,
+                completion_tokens: 5,
+                reasoning_tokens: 0,
+                cached_tokens: 0,
+                timing: Some(FinishTiming {
+                    prefill_ms: 12,
+                    decode_ms: 340,
+                    prefill_tokens: 10,
+                    phases,
+                }),
+            })
+            .await
+            .expect("send");
+            drop(tx);
+            let r = collect_chat_completion(rx, "id".into(), 1, "m".into())
+                .await
+                .expect("collected");
+            serde_json::to_value(r.usage.expect("usage").helexa_timing.expect("timing"))
+                .expect("serialise")
+        }
+
+        let measured = timing_json(Some(PhaseTiming {
+            forward_ms: 300,
+            sample_ms: 25,
+            emit_ms: 4,
+        }))
+        .await;
+        assert_eq!(measured["forward_ms"], 300);
+        assert_eq!(measured["sample_ms"], 25);
+        assert_eq!(measured["emit_ms"], 4);
+
+        let unmeasured = timing_json(None).await;
+        for f in ["forward_ms", "sample_ms", "emit_ms"] {
+            assert!(
+                unmeasured.get(f).is_none(),
+                "unmeasured {f} must be absent, not {:?} -- a zero here reads as \
+                 'this phase is free' rather than 'nobody measured it'",
+                unmeasured.get(f)
+            );
+        }
+        // The prefill/decode split is unconditional and must be unaffected.
+        assert_eq!(unmeasured["decode_ms"], 340);
+    }
+
     /// A dropped engine slot ends the channel with no Finish. Returning
     /// a 200 with partial text would hide a fatal worker fault.
     #[tokio::test]
@@ -898,6 +965,7 @@ mod tests {
                 prefill_ms: 200,
                 decode_ms: 1500,
                 prefill_tokens: 128,
+                phases: None,
             }),
         })
         .await

crates/neuron/src/wire/openai_responses.rs

@@ -219,9 +219,6 @@ pub fn request_to_chat(req: ResponsesRequest) -> Result<ChatCompletionRequest, T
         }
     }
 
-    // Only the newest turn's thinking is context; the rest is noise.
-    prune_stale_reasoning(&mut messages);
-
     // Carry the caller's extension fields across the hop (#277).
     //
     // This used to start from an empty map and insert exactly one key,
@@ -455,54 +452,6 @@ fn merge_tool_call_into(prev: &mut ChatMessage, msg: &mut ChatMessage) -> bool {
     true
 }
 
-/// Keep replayed reasoning on the most recent assistant turn only.
-///
-/// #277 restored the round-trip, and then over-corrected: a client that
-/// replays every turn verbatim — which is what an agentic harness does —
-/// had the model's entire thinking history rendered back into the
-/// prompt, because Qwen3.8's template defaults `preserve_thinking` to
-/// true. Measured on a live session: **62% of the assistant content
-/// being replayed was prior think blocks** (95,311 characters of
-/// thinking against 467 characters of actual output), and the model's
-/// per-turn reasoning collapsed to tens of tokens while it made
-/// increasingly shallow passes at the task.
-///
-/// Reasoning models are trained on histories that carry *final content*,
-/// not the deliberation behind it — Qwen's own multi-turn convention
-/// strips prior think blocks. But the newest block is different: when a
-/// turn was cut off mid-thought and the caller says "continue", that
-/// thinking is the work being resumed, which is what #277 exists to
-/// preserve.
-///
-/// So: newest assistant turn keeps its reasoning, everything older
-/// loses it. Note the template's own `preserve_thinking: false` is
-/// exactly backwards for this — it keeps thinking on turns *after* the
-/// last user message (the agentic case, where it is noise) and drops it
-/// on turns *before* (the continuation case, where it is the point).
-fn prune_stale_reasoning(messages: &mut [ChatMessage]) {
-    let newest = messages
-        .iter()
-        .rposition(|m| m.role == "assistant" && m.extra.get("reasoning_content").is_some());
-    let Some(newest) = newest else { return };
-    let mut dropped = 0usize;
-    for (i, m) in messages.iter_mut().enumerate() {
-        if i == newest || m.role != "assistant" {
-            continue;
-        }
-        if let Some(obj) = m.extra.as_object_mut()
-            && obj.remove("reasoning_content").is_some()
-        {
-            dropped += 1;
-        }
-    }
-    if dropped > 0 {
-        tracing::debug!(
-            dropped,
-            "responses: replaying reasoning for the newest assistant turn only"
-        );
-    }
-}
-
 /// The text of a replayed reasoning item, under either spelling.
 ///
 /// `summary` is what neuron emits (`summary_text` parts) and `content`
@@ -1934,14 +1883,28 @@ mod tests {
         );
     }
 
-    /// An agentic harness replays every turn verbatim. Rendering all of
-    /// that thinking back into the prompt made the context mostly
-    /// self-quotation — 62% of replayed assistant content on a live
-    /// session — and the model started making shallow passes. Only the
-    /// newest turn's reasoning is context; the rest is noise the model
-    /// was never trained to read.
+    /// Prior-turn reasoning reaches the model; the model's own chat
+    /// template decides what to do with it (#329).
+    ///
+    /// The translator used to keep reasoning on the newest assistant
+    /// turn and delete it from the rest. That was a third policy, owned
+    /// by us, that no model vendor implements: every vendor's rule is
+    /// tool-loop aware (keep thinking within the current tool-calling
+    /// episode, drop it across user turns), which is exactly what
+    /// Qwen's `last_query_index` does and what Anthropic requires
+    /// inside a tool-use turn.
+    ///
+    /// It also made the prompt non-append-only: the block the model
+    /// generated into the KV cache was deleted on the next turn, so
+    /// turn N's token sequence was no longer a prefix of turn N+1's.
+    /// Measured across the change on a live fleet, prefix-cache reuse
+    /// on this surface went 53.9% -> 0.5%.
+    ///
+    /// Policy now lives where it is expressible: the template, through
+    /// `chat_template_kwargs.preserve_thinking`, with precedence
+    /// request > operator > the template's own default.
     #[test]
-    fn only_the_newest_turns_reasoning_is_replayed() {
+    fn prior_turn_reasoning_is_left_for_the_template_to_judge() {
         let raw = r#"{
             "model": "Qwen/Qwen3.8-27B",
             "input": [
@@ -1962,10 +1925,9 @@ mod tests {
             .collect();
         assert_eq!(
             carried,
-            ["newer plan"],
-            "older turns must keep their content but lose their thinking"
+            ["old plan", "newer plan"],
+            "every replayed turn's thinking must survive translation"
         );
-        // The turns themselves survive — only the thinking is pruned.
         assert_eq!(
             chat.messages
                 .iter()
@@ -1975,8 +1937,29 @@ mod tests {
         );
     }
 
-    /// The continuation case still works: a turn cut off mid-thought is
-    /// the newest assistant turn, so its reasoning is what survives.
+    /// The caller's own policy still reaches the template unaltered —
+    /// this is the toggle #329 specifies, and the one a client can
+    /// actually set (pi sends it verbatim via `samplingParams`).
+    #[test]
+    fn a_callers_preserve_thinking_reaches_the_template_kwargs() {
+        let raw = r#"{
+            "model": "Qwen/Qwen3.8-27B",
+            "input": [{"type": "message", "role": "user", "content": "hi"}],
+            "chat_template_kwargs": {"preserve_thinking": false}
+        }"#;
+        let req: ResponsesRequest = serde_json::from_str(raw).expect("parse");
+        let chat = request_to_chat(req).expect("translate");
+        assert_eq!(
+            chat.extra
+                .get("chat_template_kwargs")
+                .and_then(|k| k.get("preserve_thinking")),
+            Some(&serde_json::Value::Bool(false))
+        );
+    }
+
+    /// The continuation case (#277): a turn cut off mid-thought is
+    /// replayed as reasoning with no assistant turn after it, and that
+    /// thinking — the work being resumed — must reach the model.
     #[test]
     fn a_truncated_turn_keeps_its_reasoning_through_a_continue() {
         let raw = r#"{

crates/neuron/tests/api.rs

@@ -99,6 +99,49 @@ async fn test_version_endpoint() {
     assert!(!body.git_sha.is_empty());
 }
 
+#[tokio::test]
+async fn test_metrics_endpoint_serves_prometheus_text() {
+    // The recorder is process-wide and other tests in this binary may
+    // have installed it already; either way the route must answer, and
+    // must never answer 200-with-empty-body (indistinguishable from a
+    // healthy process reporting nothing).
+    let _ = neuron::metrics::install();
+    neuron::metrics::record_finish("probe-model", 100, 1000, 64, Some((900, 50, 10)));
+    let url = spawn_neuron(fake_discovery()).await;
+
+    let resp = reqwest::Client::new()
+        .get(format!("{url}/metrics"))
+        .send()
+        .await
+        .expect("request should succeed");
+
+    assert_eq!(resp.status(), 200);
+    let ct = resp
+        .headers()
+        .get(reqwest::header::CONTENT_TYPE)
+        .and_then(|v| v.to_str().ok())
+        .unwrap_or_default()
+        .to_string();
+    assert!(
+        ct.starts_with("text/plain"),
+        "Prometheus needs a text/plain scrape, got {ct:?}"
+    );
+
+    let body = resp.text().await.expect("body");
+    // The exporter renders only metrics that carry samples -- a
+    // `describe_*` on its own produces nothing -- so the scrape has to
+    // be checked after a recorded request, which is also the path that
+    // actually matters: record here, readable there.
+    assert!(
+        body.contains("neuron_decode_phase_seconds"),
+        "a recorded request must appear in the scrape:\n{body}"
+    );
+    assert!(
+        body.contains(r#"phase="forward""#) && body.contains(r#"phase="residual""#),
+        "phase labels must survive to the scrape:\n{body}"
+    );
+}
+
 #[tokio::test]
 async fn test_health_endpoint() {
     let url = spawn_neuron(fake_discovery()).await;

helexa-router.toml

@@ -28,6 +28,9 @@ region = "eu-central"
 [aliases]
 "helexa/small" = "Qwen/Qwen3-8B"
 "helexa/balanced" = "Qwen/Qwen3.8-27B"
+# System-One decisions (#333). Also the model /v1/systemone falls back
+# to when a request names none or names a service no operator serves.
+"helexa/one" = "convaiinnovations/laya"
 
 # hanzalova operator cortex, plaintext over the WireGuard mesh (the mesh
 # is the trust boundary; no TLS pin needed for http://).

helexa.ai/content/docs/2-operating/03-neurons.md

@@ -358,6 +358,20 @@ Whether narrowing replay *helps* the model is unmeasured. Set it in both
 `/v1/models` and logged at load so a session can be attributed to the
 arm it ran under.
 
+A caller can override it per request with
+`chat_template_kwargs.preserve_thinking`, on `/v1/chat/completions` and
+`/v1/responses` alike — precedence is request > operator > template
+default.
+
+The setting also decides whether a multi-turn prompt is append-only, and
+therefore whether the prefix cache can work at all. Replay is
+byte-stable: the think block the model generated is re-rendered
+unchanged, so turn N's tokens stay a prefix of turn N+1's. Anything that
+deletes a prior turn's reasoning rewrites the middle of the prompt and
+invalidates the cache from that point. Measured on this fleet when the
+Responses path pruned prior reasoning unconditionally, reuse fell from
+53.9% to 0.5% and stayed there (#329).
+
 **`reasoning_budget.answer_reserve_tokens`** (default `4096`) holds back
 part of the caller's output budget for the answer:
 

models.toml

@@ -167,6 +167,28 @@ min_device_vram_mb = 4000
 cost.input = 0.015
 cost.output = 0.06
 
+# System-One decisions (#333): Laya, a 421M non-autoregressive decision
+# model (ModernBERT encoder + option-scoring head) served on
+# /v1/systemone. One repo, three checkpoints -- English at the root,
+# multilingual/ and typed-decisions/ beside it -- all resident (~2.3 GB
+# bf16) so neuron can route each request by language. Not a text model:
+# chat endpoints answer it with wrong_modality.
+#
+# Resident on quadbrat, replacing the Qwen3-1.7B pre-warm (operator
+# decision 2026-09-29, #341); pinned there so a cold-loaded helexa/small
+# can share the card but never evict it. Metered like TypeSafe Jev,
+# input tokens only (every question re-encodes the state, and each
+# encoded token counts), at Jev's published $0.042 per million.
+[[models]]
+id = "convaiinnovations/laya"
+harness = "candle"
+vram_mb = 3500
+min_devices = 1
+min_device_vram_mb = 4000
+pinned_on = ["quadbrat"]
+cost.input = 0.042
+cost.output = 0.0
+
 # -- Tier aliases ------------------------------------------------------------
 # Clients can request inference against an alias and cortex transparently
 # routes to the concrete model id below. Useful so client code can pick a
@@ -177,6 +199,10 @@ cost.output = 0.06
 "helexa/balanced" = "Qwen/Qwen3-8B"
 "helexa/large" = "Qwen/Qwen3.8-27B"
 "helexa/image" = "Tongyi-MAI/Z-Image-Turbo"
+# The System-One decision class: the default /v1/systemone resolves to
+# for a request with no model, or one naming a service this fleet does
+# not serve (a Jev client's own model name).
+"helexa/one" = "convaiinnovations/laya"
 
 # Text-to-image (#197/#203): Z-Image-Turbo, 6B S3-DiT + Qwen3-4B text
 # encoder (runs on host CPU by default — see [harness.candle.image] in

script/generate_laya_fixture.py

@@ -0,0 +1,160 @@
+"""Generate a tiny Laya parity fixture from the upstream reference (#336).
+
+Builds the `laya` SDK's own `DecisionModel` (laya/common.py) around a
+tiny ModernBERT, then records its outputs for a right-padded batch.
+
+Writes, into --out:
+  rl_agent_config.json   the decision-model config, as neuron parses it
+  encoder/config.json    the ModernBERT config, saved by transformers
+  model.safetensors      the reference's own weights, float32
+  tokenizer.json         a placeholder word-level tokenizer (the fixture
+                         is scored on token ids; loading needs a file)
+  expected.json          the batch (ids, markers, question types) and the
+                         reference's per-row scorer logits and act softmax
+
+The shape is chosen so every mechanism is exercised at toy size:
+
+- four encoder layers with a global layer on each side of two sliding
+  ones, and a sliding window (local_attention = 8, so +-4 positions)
+  far shorter than the 40-token row, so a window bug changes the result;
+- two different rotary bases for global and sliding layers, so using
+  the wrong one for either kind is visible;
+- rows of different lengths, so padding is present and must be masked;
+- a one-option row, which takes the reference's single-marker branch
+  of the act features;
+- every question type, so the type embedding is indexed three ways.
+
+Weights are initialised well above transformers' default scale so
+attention is sharp rather than near-uniform, and norm weights and biases
+are perturbed away from their ones/zeros init: at init a missing norm
+and a present one are indistinguishable.
+
+Usage, in a venv holding the `laya` SDK and CPU torch:
+
+    python script/generate_laya_fixture.py \
+        --out crates/neuron/src/harness/testdata/laya/tiny
+"""
+
+import argparse
+import json
+import os
+
+import torch
+from safetensors.torch import save_file
+from transformers import ModernBertConfig, AutoModel
+
+from laya.common import DecisionModel
+
+ap = argparse.ArgumentParser()
+ap.add_argument("--out", required=True)
+ap.add_argument("--seed", type=int, default=7)
+args = ap.parse_args()
+torch.manual_seed(args.seed)
+os.makedirs(os.path.join(args.out, "encoder"), exist_ok=True)
+
+ecfg = ModernBertConfig(
+    vocab_size=97,
+    hidden_size=32,
+    num_hidden_layers=4,
+    num_attention_heads=2,
+    intermediate_size=48,
+    max_position_embeddings=128,
+    local_attention=8,
+    global_attn_every_n_layers=3,
+    pad_token_id=0,
+    bos_token_id=1,
+    eos_token_id=2,
+    cls_token_id=1,
+    sep_token_id=2,
+    norm_eps=1e-5,
+    hidden_activation="gelu",
+    # Well above the 0.02 default: at 0.02 attention is near-uniform and
+    # a window or rotary bug moves the logits only in the sixth decimal.
+    initializer_range=0.3,
+    rope_parameters={
+        "full_attention": {"rope_theta": 1000.0, "rope_type": "default"},
+        "sliding_attention": {"rope_theta": 300.0, "rope_type": "default"},
+    },
+)
+encoder = AutoModel.from_config(ecfg, attn_implementation="sdpa")
+model = DecisionModel(encoder, head_layers=2, n_act=2).eval()
+
+with torch.no_grad():
+    for name, p in model.named_parameters():
+        if "norm" in name or name.startswith("scorer.0"):
+            p.add_(0.3 * torch.randn_like(p))
+        elif name.startswith("type_emb"):
+            p.mul_(3.0)
+
+ecfg.save_pretrained(os.path.join(args.out, "encoder"))
+rl_cfg = {
+    "encoder": "tiny-modernbert",
+    "head_layers": 2,
+    "max_len": 64,
+    "head_max_len": 24,
+    "act_costs": {"escalate": 0.5},
+    "amp_dtype": "bf16",
+    "temperature": [1.0, 1.0, 1.0],
+    "temperature_by_options": {},
+}
+with open(os.path.join(args.out, "rl_agent_config.json"), "w") as f:
+    json.dump(rl_cfg, f, indent=1)
+# A placeholder: the fixture is scored on token ids, but a checkpoint
+# directory is loaded with its tokenizer, so one must exist. It carries
+# the special tokens a decision tokenizer is looked up by, with [PAD] at
+# the encoder's pad_token_id (0).
+with open(os.path.join(args.out, "tokenizer.json"), "w") as f:
+    json.dump({"version": "1.0", "truncation": None, "padding": None, "added_tokens": [],
+               "normalizer": None, "pre_tokenizer": {"type": "Whitespace"},
+               "post_processor": None, "decoder": None,
+               "model": {"type": "WordLevel",
+                         "vocab": {"[PAD]": 0, "[UNK]": 1, "[CLS]": 2, "[SEP]": 3, "[MASK]": 4,
+                                   "a": 5, "b": 6},
+                         "unk_token": "[UNK]"}}, f)
+with open(os.path.join(args.out, "tokenizer_config.json"), "w") as f:
+    json.dump({"cls_token": "[CLS]", "sep_token": "[SEP]", "mask_token": "[MASK]",
+               "pad_token": "[PAD]", "unk_token": "[UNK]"}, f, indent=1)
+
+state = {k: v.detach().contiguous().float() for k, v in model.state_dict().items()}
+save_file(state, os.path.join(args.out, "model.safetensors"))
+
+# (length, marker positions, qtype). Ids avoid 0 (pad) inside a row.
+rows = [
+    (40, [3, 9, 17], 0),
+    (23, [4, 11], 1),
+    (9, [2], 2),
+    (31, [5, 7, 12, 20, 26], 0),
+]
+seqs = []
+for n, markers, qtype in rows:
+    ids = torch.randint(1, ecfg.vocab_size, (n,)).tolist()
+    seqs.append({"input_ids": ids, "markers": markers, "qtype": qtype})
+
+b = len(seqs)
+L = max(len(s["input_ids"]) for s in seqs)
+K = max(len(s["markers"]) for s in seqs)
+ids = torch.zeros((b, L), dtype=torch.long)
+att = torch.zeros((b, L), dtype=torch.long)
+mpos = torch.zeros((b, K), dtype=torch.long)
+mmask = torch.zeros((b, K), dtype=torch.bool)
+for i, s in enumerate(seqs):
+    n = len(s["input_ids"])
+    ids[i, :n] = torch.tensor(s["input_ids"])
+    att[i, :n] = 1
+    k = len(s["markers"])
+    mpos[i, :k] = torch.tensor(s["markers"])
+    mmask[i, :k] = True
+qtype = torch.tensor([s["qtype"] for s in seqs])
+
+with torch.no_grad():
+    logits, act = model(ids, att, mpos, mmask, qtype)
+act = torch.softmax(act.float(), -1)
+
+for i, s in enumerate(seqs):
+    k = len(s["markers"])
+    s["logits"] = logits[i, :k].tolist()
+    s["act"] = act[i].tolist()
+
+with open(os.path.join(args.out, "expected.json"), "w") as f:
+    json.dump({"seed": args.seed, "seqs": seqs}, f, indent=1)
+print("wrote", args.out, "params:", sum(v.numel() for v in state.values()))

script/laya-lang-reference.py

@@ -0,0 +1,361 @@
+#!/usr/bin/env python3
+"""Record Laya's language/script detector over a broad corpus (#339).
+
+neuron routes each decision request to the English or multilingual Laya
+checkpoint with a port of `laya.lang.analyse`. That function is a pile of
+heuristics sitting on Python's Unicode semantics -- `str.isalpha`, `re`'s
+`\\w`, `unicodedata.combining`, `str.lower`, `str.split` -- and a port that
+reads the same can still route differently. This records the reference's
+answer over a corpus built to reach every rule and every place those
+semantics could diverge:
+
+  - every script range, real sentences and generated letter runs
+  - each function-word language, accented and ASCII-stripped, long and short
+  - English with loanwords, names, symbols and IPA; code, URLs, e-mails,
+    identifiers, acronyms and slash compounds
+  - mixed-script and mixed-language states, multi-line and structured
+  - the 4000-character budgets, the nesting depth limit, non-string leaves
+  - combining marks, decomposed text, case-mapping oddities (İ, ß, final
+    sigma, titlecase digraphs), fullwidth forms, numerals, emoji, and the
+    whitespace Python and Rust disagree about
+  - a seeded fuzz section mixing all of the above
+
+Every mapping in the corpus has its keys in sorted order, so a consumer
+whose JSON maps iterate in key order (serde_json without `preserve_order`)
+sees the same leaf order Python does.
+
+Characters are limited to those assigned in the Unicode version of the
+Python that records the corpus, so a newer Unicode table on the consuming
+side cannot disagree about a code point Python considers unassigned.
+
+Usage (any Python 3.10+ with the laya SDK importable; no model, no network):
+
+    python script/laya-lang-reference.py \\
+        --out crates/neuron/src/harness/testdata/laya/lang_corpus.json
+"""
+import argparse
+import json
+import random
+import sys
+import unicodedata
+
+import laya
+from laya.lang import _WORD, _iter_text, analyse
+
+SENTENCES = {
+    "en": "Please refund the duplicate charge, I was billed twice this month and it is not fair.",
+    "en_short": "refund please",
+    "en_loanword": "I would like a café table and my résumé is attached for the role.",
+    "en_name": "José and Zoë have the tickets that were sent to you from the office.",
+    "en_two_loanwords": "The naïve café owner sent the résumé to you and it was not what we wanted.",
+    "en_symbols": "Set α to 0.05 and β to 0.9, then run the job again with the new values.",
+    "en_russian_name": "The report from Дмитрий Петрович Савицкий was sent to the team yesterday.",
+    "en_ipa": "Vladimir is pronounced [vlɐˈdʲimʲɪr] in Russian and that is how we say it.",
+    "en_stress": "The name Влади́мир has a stress mark on the second syllable for you.",
+    "fr": "Bonjour, je voudrais annuler ma commande car elle est arrivée très en retard.",
+    "fr_short": "merci beaucoup",
+    "de": "Die Rechnung ist falsch, bitte korrigieren Sie den Betrag für diesen Monat.",
+    "de_short": "Danke schön",
+    "de_eszett": "Ich möchte die Straße und die Größe meiner Bestellung ändern, bitte.",
+    "es": "Me han cobrado dos veces este mes y quiero que me devuelvan el dinero ya.",
+    "pt": "Não consigo acessar a minha conta, já tentei redefinir a senha várias vezes.",
+    "pt_br": "Voce pode me mandar a nota fiscal? Nao recebi nada ate agora.",
+    "pt_ticket": "Deu erro 500 no endpoint de login depois do update de ontem.",
+    "it": "Vorrei sapere quando arriverà il mio ordine, è già in ritardo di una settimana.",
+    "nl": "Het pakket is niet aangekomen en ik wil graag een terugbetaling voor deze bestelling.",
+    "ro": "Vreau să anulez comanda pentru că produsul este deteriorat și nu funcționează.",
+    "bn": "ami taka ferot chai, amar order ekhono ashe nai, ki korbo bolun",
+    "az": "Mən sifarişi ləğv etmək istəyirəm, çünki məhsul çox gec gəldi.",
+    "tr": "Siparişimi iptal etmek istiyorum çünkü ürün çok geç geldi ve kırık.",
+    "pl": "Chciałbym zwrócić pieniądze, ponieważ paczka dotarła uszkodzona.",
+    "cs": "Chtěl bych vrátit peníze, protože zásilka dorazila poškozená.",
+    "hu": "Szeretném visszakapni a pénzemet, mert a csomag sérülten érkezett.",
+    "vi": "Tôi muốn được hoàn tiền vì gói hàng đã đến bị hư hỏng.",
+    "sv": "Jag vill ha pengarna tillbaka eftersom paketet kom fram skadat.",
+    "da": "Jeg vil gerne have pengene tilbage, fordi pakken kom i stykker.",
+    "sw": "Nataka kurudishiwa pesa zangu kwa sababu kifurushi kiliharibika.",
+    "la_shared_only": "la la e e o o un un",
+    "el": "Θέλω να ακυρώσω την παραγγελία μου γιατί ήρθε πολύ αργά.",
+    "el_upper_sigma": "ΘΕΛΩ ΤΗΝ ΕΠΙΣΤΡΟΦΗ ΧΡΗΜΑΤΩΝ ΤΟΥ ΛΟΓΑΡΙΑΣΜΟΥ ΣΑΣ",
+    "ru": "Мне дважды списали деньги в этом месяце, верните, пожалуйста.",
+    "uk": "Мені двічі списали гроші цього місяця, поверніть, будь ласка.",
+    "hy": "Ես ուզում եմ չեղարկել պատվերս, քանի որ այն ուշ հասավ։",
+    "he": "אני רוצה לבטל את ההזמנה שלי כי היא הגיעה באיחור.",
+    "he_niqqud": "שָׁלוֹם, אֲנִי רוֹצֶה לְבַטֵּל אֶת הַהַזְמָנָה",
+    "ar": "تم خصم المبلغ مرتين هذا الشهر، أريد استرداد أموالي من فضلكم.",
+    "fa": "می‌خواهم سفارشم را لغو کنم چون خیلی دیر رسید.",
+    "hi": "मुझे इस महीने दो बार शुल्क लिया गया, मुझे पैसे वापस चाहिए।",
+    "bn_script": "এই মাসে আমার কাছ থেকে দুবার টাকা কাটা হয়েছে, টাকা ফেরত চাই।",
+    "pa": "ਮੈਨੂੰ ਇਸ ਮਹੀਨੇ ਦੋ ਵਾਰ ਚਾਰਜ ਕੀਤਾ ਗਿਆ, ਪੈਸੇ ਵਾਪਸ ਚਾਹੀਦੇ ਹਨ।",
+    "gu": "મને આ મહિને બે વાર ચાર્જ કરવામાં આવ્યો, પૈસા પાછા જોઈએ છે.",
+    "or": "ମୋତେ ଏହି ମାସରେ ଦୁଇଥର ଚାର୍ଜ କରାଯାଇଛି।",
+    "ta": "இந்த மாதம் எனக்கு இரண்டு முறை கட்டணம் வசூலிக்கப்பட்டது.",
+    "te": "ఈ నెలలో నాకు రెండుసార్లు ఛార్జ్ చేశారు, డబ్బు తిరిగి ఇవ్వండి.",
+    "kn": "ಈ ತಿಂಗಳು ನನಗೆ ಎರಡು ಬಾರಿ ಶುಲ್ಕ ವಿಧಿಸಲಾಗಿದೆ.",
+    "ml": "ഈ മാസം എനിക്ക് രണ്ടുതവണ ചാർജ് ഈടാക്കി.",
+    "si": "මෙම මාසයේ මට දෙවරක් ගාස්තු අය කළා.",
+    "th": "เดือนนี้ฉันถูกเรียกเก็บเงินสองครั้ง ขอเงินคืนด้วยค่ะ",
+    "lo": "ເດືອນນີ້ຂ້ອຍຖືກເກັບເງິນສອງເທື່ອ",
+    "bo": "ང་ལ་ཟླ་བ་འདིར་ཐེངས་གཉིས་རིན་བསྡུས་སོང་།",
+    "my": "ဒီလမှာ ကျွန်တော့်ကို နှစ်ကြိမ် ငွေတောင်းခံခဲ့တယ်။",
+    "ka": "ამ თვეში ორჯერ ჩამომეჭრა თანხა, გთხოვთ დამიბრუნოთ.",
+    "am": "በዚህ ወር ሁለት ጊዜ ክፍያ ተቀንሶብኛል፣ እባክዎ ይመልሱልኝ።",
+    "km": "ខ្ញុំត្រូវបានគិតប្រាក់ពីរដងក្នុងខែនេះ ខ្ញុំចង់បានប្រាក់វិញ",
+    "ko": "이번 달에 두 번 청구되었습니다. 환불해 주세요.",
+    "ja": "今月二回請求されました。返金をお願いします。",
+    "ja_kana": "ありがとうございます。キャンセルしたいです。",
+    "zh": "这个月我被扣了两次钱，我要退款，请尽快处理。",
+    "zh_trad": "這個月我被扣了兩次錢，我要退款。",
+    "yue_ext_b": "𠮷野家の𩸽を食べたい",
+    "mn_cyr": "Би захиалгаа цуцлахыг хүсч байна.",
+    "bopomofo": "ㄅㄆㄇㄈ ㄉㄊㄋㄌ",
+    "halfwidth_kana": "ｷｬﾝｾﾙｼﾀｲﾃﾞｽ",
+    "cherokee": "ᏣᎳᎩ ᎦᏬᏂᎯᏍᏗ",
+}
+
+EXTRA = [
+    "", " ", "   \n\t  ", "12345 67890", "3.14 2.71 1.41", "!!! ??? ...", "ok", "a", "no",
+    "🙂🙂🙂 👍", "refund 🙂 please 👍 thanks", "café", "résumé naïve café",
+    "ß", "Straße", "İstanbul İzmir", "ıi İI", "Džemal Ljubljana Nj", "ⓐⓑⓒ ⒶⒷⒸ", "𝐀𝐁𝐂 𝐚𝐛𝐜 test",
+    "Ⅰ Ⅱ Ⅲ ⅳ ⅴ", "x² y³ ½ ¾", "٣٤٥ ١٢", "१२३ ४५६", "ＦＵＬＬＷＩＤＴＨ ｔｅｘｔ ｈｅｒｅ",
+    "ʰʲʷ ˈˌ", "ɐɑɒ ʃʒ θð", "ƀƁƂ ǝǞ ȀȁȂ", "Ȿ ɀ", "ḀḁḂḃ ẞ ỳỹ",
+    "a\x1cb\x1dc\x1ed\x1ff", "word\u00a0word\u00a0word\u00a0word", "一\u3000二\u3000三",
+    "tab\tseparated\tvalues\there", "zero\u200bwidth\u200bjoiner", "soft\u00adhyphen",
+    "e\u0301 e\u0300 a\u0308 o\u0303", "Tôi muốn được hoàn tiền",
+    unicodedata.normalize("NFD", "Tôi muốn được hoàn tiền vì gói hàng đã đến bị hư hỏng."),
+    unicodedata.normalize("NFD", "Não consigo acessar a minha conta, já tentei várias vezes."),
+    unicodedata.normalize("NFD", "Vreau să anulez comanda pentru că produsul este deteriorat."),
+    "ΟΔΟΣ ΟΔΟΣ ΟΔΟΣ", "Σ", "ΑΣ ΑΣ", "ὈΔΥΣΣΕΎΣ",
+]
+
+CODE_AND_IDENTIFIERS = [
+    "see https://github.com/acme/os.path and user@acme.com for details",
+    "email me at joao.silva@empresa.com.br or maria@exemplo.pt please",
+    "version v1.2.3 of the lib, see U.S.A. rules and e.g. the docs",
+    "x = round(el, 2); return non_english",
+    "import os.path\nfrom foo import bar",
+    "def f(x):\n    return {'la': 1, 'le': 2}",
+    "Nav/Com OS/2 C:\\DOS\\mode ESA/UN are names not sentences at all",
+    "MON LA EST COM DES are teams, states and radio bands in the listing",
+    "COM COM COM COM radio listing for the English region",
+    "O CLIENTE ESTÁ MUITO IRRITADO COM O ATRASO DO PEDIDO",
+    "call foo(bar) then baz(qux) and see what the output of the call is",
+    "Deu erro (500) no login depois do update de ontem",
+    "a.b.c.d e@f g.h-i j_k.l m-n.o",
+    "-leading.dash and trailing.dot. and @handle and .hidden files here",
+    "..... @@@@ .@. a.@b",
+    "x" * 300 + ".com",
+    "path/to/file.py line 42 in <module> raise ValueError('não funciona')",
+]
+
+MULTILINE = [
+    "Hello team,\nPlease see below.\nNão consigo acessar a minha conta, já tentei redefinir a senha.\nThanks",
+    "Traceback (most recent call last):\n  File \"app.py\", line 3, in <module>\n    x = foo()\nO sistema não "
+    "está funcionando desde ontem e preciso de ajuda urgente com isso",
+    "Line one in English is here.\nLa fattura è sbagliata, per favore correggetela subito.",
+    "Short line\n\n\nAnother short\nDie Rechnung ist falsch und ich bitte um eine Korrektur heute.",
+    "English only line one.\nEnglish only line two.\nAnd the third line of English text.",
+    "José\nMaría\nFrançois\nplease refund the order that was sent to them",
+    "Ticket #1\nこんにちは、注文をキャンセルしたいです\nThanks for the help with this",
+    # Targeted: each of the next cases flips if one rule is ported wrong.
+    # All-caps function words in mixed-case text are acronyms, not French.
+    "Please refund the duplicate charge, I was billed twice this month and it is not fair.\n"
+    "Scores: MON LA EST DES COM LES ET tonight",
+    # U+001C splits tokens for str.split(), so the identifier is dropped
+    # alone rather than taking the Portuguese words with it.
+    "Please refund the duplicate charge, I was billed twice this month.\n"
+    "Não\x1cconsigo\x1cacessar\x1cminha\x1cconta\x1cjá\x1cx.y",
+    # A combining stress mark must not split a capitalised name into a
+    # lowercase "word".
+    "The customer name on the invoice is Влади́мир "
+    "Петро́вич and we need it fixed today",
+    # One loanword: English function words rescue the marginal diacritic rate.
+    "We will meet at the café for you",
+    "I think the naïve plan is fine for all of us here",
+]
+
+
+def budget_cases():
+    en = "Please refund the duplicate charge on my card for this month. "
+    pt = "Não consigo acessar a minha conta, já tentei redefinir a senha várias vezes. "
+    out = []
+    for n in (3990, 3999, 4000, 4001, 4010, 8000):
+        out.append((en * (n // len(en) + 1))[:n])
+    out.append((en * 70)[:3995] + " " + pt)
+    out.append((en * 70) + pt)
+    out.append((en * 100) + "\n" + pt)
+    out.append("x" * 3999 + "\n" + pt)
+    out.append("x" * 4000 + "\n" + pt)
+    out.append(("é" * 5000))
+    out.append(("ab " * 1400) + pt)
+    return out
+
+
+def structured_cases():
+    en = SENTENCES["en"]
+    pt = SENTENCES["pt"]
+    de = SENTENCES["de"]
+    ja = SENTENCES["ja"]
+    deep = "Não consigo acessar a minha conta, já tentei redefinir a senha."
+    nested = deep
+    deeps = []
+    for depth in range(9):
+        nested = [nested]
+        deeps.append(nested)
+    return [
+        {"body": pt, "subject": "Refund request"},
+        {"body": en, "note": pt},
+        {"a_note": en * 60, "b_message": de},
+        {"a_note": en * 60, "b_message": "Hallo, bitte."},
+        {"customer": "José", "message": en},
+        {"name": "Дмитрий Петрович", "text": en},
+        {"message": ja, "subject": "Order"},
+        {"count": 3, "flag": True, "missing": None, "text": en},
+        {"count": 3, "flag": False, "text": pt},
+        {"items": [en, pt, {"deep": de}], "meta": {"lang": "?"}},
+        {"conversation": [{"content": en, "role": "user"}, {"content": pt, "role": "user"}]},
+        {"a": {"b": {"c": {"d": {"e": {"f": {"g": pt}}}}}}},
+        {"a": {"b": {"c": {"d": {"e": {"f": pt}}}}}},
+        {"code": "x = round(el, 2)\nreturn os.path", "text": en},
+        {"body": "Traceback:\n  File x.py\nO sistema não está funcionando desde ontem e preciso de ajuda", "id": 7},
+        {"a": "José", "b": "María", "c": "François", "d": en},
+        {"a": en * 70, "b": "Vreau să anulez comanda pentru că produsul este deteriorat."},
+        {"a": en * 70, "b": "Мне дважды списали деньги в этом месяце, верните."},
+        {"a": "ok", "b": "no", "c": "yes"},
+        {},
+        [],
+        [en, pt],
+        [pt, en],
+        [en, en, en, "Die Rechnung ist falsch, bitte korrigieren."],
+        ["hello", "bonjour", "hola"],
+        [1, 2, 3],
+        [None, True, 2.5],
+        [[["deep", ["deeper", [pt]]]]],
+        *deeps,
+        12345,
+        3.5,
+        True,
+        None,
+    ]
+
+
+def fuzz_cases(n, seed):
+    rng = random.Random(seed)
+    pools = [v for v in SENTENCES.values()] + EXTRA + CODE_AND_IDENTIFIERS
+    words = []
+    for s in pools:
+        words.extend(s.split())
+    specials = list("\n\t .,;:=()[]{}/\\@_-'\"") + ["\x1c", "\u00a0", "\u3000", "\u0301", "\u0308",
+                                                   "\u200b", "İ", "ı", "ß", "Σ", "Dž", "Ⅲ", "²",
+                                                   "🙂", "１", "Ａ", "ａ", "ʰ", "ɐ", "é", "ñ"]
+    out = []
+    for _ in range(n):
+        k = rng.randint(1, 40)
+        parts = []
+        for _ in range(k):
+            r = rng.random()
+            if r < 0.75:
+                parts.append(rng.choice(words))
+            else:
+                parts.append(rng.choice(specials))
+        sep = rng.choice([" ", " ", " ", "\n", ""])
+        out.append(sep.join(parts))
+    return out
+
+
+def generated_script_runs():
+    """A run of assigned letters from every script range the detector names."""
+    from laya.lang import _SCRIPT_RANGES
+    out = []
+    for name, ranges in _SCRIPT_RANGES:
+        for lo, hi in ranges:
+            letters = [chr(cp) for cp in range(lo, hi + 1)
+                       if unicodedata.category(chr(cp)) != "Cn"]
+            if not letters:
+                continue
+            step = max(1, len(letters) // 24)
+            sample = letters[::step][:24]
+            out.append("".join(sample))
+            out.append(" ".join("".join(sample[i:i + 4]) for i in range(0, len(sample), 4)))
+    return out
+
+
+def assigned(s):
+    return all(unicodedata.category(c) != "Cn" and not (0xD800 <= ord(c) <= 0xDFFF) for c in s)
+
+
+def sorted_keys(v):
+    if isinstance(v, dict):
+        return list(v) == sorted(v) and all(sorted_keys(x) for x in v.values())
+    if isinstance(v, list):
+        return all(sorted_keys(x) for x in v)
+    return True
+
+
+def all_strings(v):
+    if isinstance(v, str):
+        yield v
+    elif isinstance(v, dict):
+        for x in v.values():
+            yield from all_strings(x)
+    elif isinstance(v, list):
+        for x in v:
+            yield from all_strings(x)
+
+
+def main():
+    ap = argparse.ArgumentParser()
+    ap.add_argument("--out", required=True)
+    ap.add_argument("--fuzz", type=int, default=400)
+    ap.add_argument("--seed", type=int, default=339)
+    args = ap.parse_args()
+
+    states = []
+    for name, s in SENTENCES.items():
+        states.append(s)
+        ascii_s = "".join(c for c in unicodedata.normalize("NFKD", s) if not unicodedata.combining(c))
+        if ascii_s != s:
+            states.append(ascii_s)
+        states.append(s.upper())
+        states.append(" ".join(s.split()[:3]))
+    states += EXTRA + CODE_AND_IDENTIFIERS + MULTILINE + budget_cases()
+    states += generated_script_runs()
+    states += structured_cases()
+    # Pairwise mixes: one English and one foreign line, both orders.
+    for name, s in SENTENCES.items():
+        if not name.startswith("en"):
+            states.append(SENTENCES["en"] + "\n" + s)
+            states.append(s + " " + SENTENCES["en"])
+    states += fuzz_cases(args.fuzz, args.seed)
+
+    cases, seen = [], set()
+    for st in states:
+        key = json.dumps(st, ensure_ascii=False, sort_keys=False)
+        if key in seen:
+            continue
+        seen.add(key)
+        assert sorted_keys(st), "mapping keys must be sorted: %r" % (st,)
+        if not all(assigned(s) for s in all_strings(st)):
+            continue
+        det = analyse(st)
+        # Recorded separately because a JSON reader that does not keep key
+        # order loses the profile's order, which the reference defines.
+        rec = {"state": st, "leaves": _iter_text(st), "analysis": det,
+               "profile_order": list(det["script_profile"])}
+        if isinstance(st, str):
+            rec["words"] = _WORD.findall(st)
+        cases.append(rec)
+
+    header = {
+        "laya_version": laya.__version__,
+        "python": sys.version.split()[0],
+        "unidata_version": unicodedata.unidata_version,
+        "fuzz_seed": args.seed,
+    }
+    with open(args.out, "w") as f:
+        json.dump({"header": header, "cases": cases}, f, ensure_ascii=False, indent=0)
+    print("%d cases" % len(cases), file=sys.stderr)
+
+
+if __name__ == "__main__":
+    main()

script/laya-reference.py

@@ -0,0 +1,385 @@
+#!/usr/bin/env python3
+"""Record Laya reference outputs for neuron's parity tests (#334).
+
+Runs the upstream `laya` SDK (the reference implementation, not the
+`rl_common.py` shipped in the model repo, which has diverged) over a fixed
+corpus and writes, per case:
+
+  - the request exactly as a `/v1/systemone` client would send it
+  - per question: `input_ids`, `[MASK]` marker positions, question type and
+    the option-collapse stats, straight out of `Agent._encode_state`
+  - the raw scorer logits and act probabilities out of `Agent._forward`,
+    before any temperature is applied
+  - the final Jev-shaped response (`answers`, `usage`, `routing`)
+
+A separate section records `laya.lang.analyse` on a language corpus, for
+the router port.
+
+The reference runs on CPU in float32 so the numbers are deterministic and
+free of autocast noise; neuron's reduced-precision paths are compared
+against them with a tolerance.
+
+Usage (in a venv with `pip install laya==<pinned>` and CPU torch):
+
+    HF_ENDPOINT=... HF_TOKEN=... python script/laya-reference.py \
+        --out crates/neuron/src/harness/testdata/laya
+
+Point HF_ENDPOINT at whichever Hugging Face-compatible registry you fetch
+weights through. Python's huggingface_hub may additionally need
+SSL_CERT_FILE (for a private CA) and HF_HUB_DISABLE_XET=1 (for registries
+without Xet support).
+"""
+import argparse
+import json
+import os
+import sys
+
+import laya
+import numpy as np
+import torch
+from laya import Router
+from laya.agent import Agent
+from laya.lang import analyse
+from laya.serve import _resolve_model
+
+LONG_TEXT = (
+    "Hello team, following up on the invoice discrepancy from last quarter. "
+    "The amounts on INV-2291 and INV-2307 do not match the purchase order, "
+    "and our finance department needs a corrected copy before month end. "
+) * 12
+
+CASES = [
+    # --- the three primitives, default budgets
+    {
+        "name": "choice_basic",
+        "state": "I was charged twice this month, I want my money back",
+        "questions": {
+            "queue": {"type": "choice", "instructions": "Which team?",
+                      "criteria": {"billing": "billing and refunds", "tech": "login and app issues",
+                                   "other": "everything else"}},
+        },
+    },
+    {
+        "name": "score_basic",
+        "state": "This is the THIRD time I've asked. Fix it today or I'm cancelling.",
+        "questions": {
+            "urgency": {"type": "score", "instructions": "How urgent?",
+                        "criteria": ["calm", "firm", "angry", "furious"]},
+        },
+    },
+    {
+        "name": "noul_default_criteria",
+        "state": "Please refund the duplicate charge on my card.",
+        "questions": {"refund": {"type": "noul", "instructions": "Is the customer asking for a refund?"}},
+    },
+    {
+        "name": "noul_custom_criteria_and_labels",
+        "state": "Can you reset my password? I'm locked out.",
+        "questions": {
+            "locked": {"type": "noul", "instructions": "Is the user locked out of their account?",
+                       "criteria": {"true": "the user cannot sign in", "false": "the user can sign in"},
+                       "labels": {"true": "locked", "false": "not locked"}},
+        },
+    },
+    # --- several questions in one request: the shared-state, one-forward path
+    {
+        "name": "multi_question",
+        "state": "I was charged twice this month, I want my money back",
+        "questions": {
+            "queue": {"type": "choice", "instructions": "Which team?",
+                      "criteria": {"billing": "billing and refunds", "tech": "login and app issues",
+                                   "other": "everything else"}},
+            "urgency": {"type": "score", "instructions": "How urgent?",
+                        "criteria": ["calm", "firm", "angry", "furious"]},
+            "refund": {"type": "noul", "instructions": "Is the customer asking for a refund?"},
+            "tone": {"type": "choice", "instructions": "Tone?", "criteria": ["polite", "rude", "neutral"]},
+            "spam": {"type": "noul", "instructions": "Is this spam?"},
+        },
+    },
+    # --- every temperature bucket
+    {
+        "name": "choice_2",
+        "state": "The app crashes when I open settings.",
+        "questions": {"kind": {"type": "choice", "instructions": "Bug or feature request?",
+                               "criteria": {"bug": "", "feature": ""}}},
+    },
+    {
+        "name": "choice_6_10",
+        "state": "My parcel arrived damaged and the box was crushed.",
+        "questions": {"topic": {"type": "choice", "instructions": "Topic?",
+                                "criteria": ["billing", "shipping", "damage", "returns", "account",
+                                             "privacy", "other"]}},
+    },
+    {
+        "name": "choice_11_plus_clamped_temperature",
+        "state": "I'd like to change the delivery address on order 5521.",
+        "questions": {"intent": {"type": "choice", "instructions": "Intent?",
+                                 "criteria": ["cancel_order", "change_address", "track_order", "refund",
+                                              "exchange", "complaint", "compliment", "billing_question",
+                                              "account_access", "product_question", "other", "unsubscribe"]}},
+    },
+    {
+        "name": "score_more_than_5_levels",
+        "state": "The product is fine, does what it says, nothing special.",
+        "questions": {"rating": {"type": "score", "instructions": "Star rating?",
+                                 "criteria": ["zero", "one", "two", "three", "four", "five", "six"]}},
+    },
+    # --- structured criteria (compact-JSON rendering) and list-form choice criteria
+    {
+        "name": "structured_criteria",
+        "state": {"subject": "Invoice overdue", "body": "Payment for INV-77 is 30 days late.", "priority": 2},
+        "questions": {
+            "queue": {"type": "choice", "instructions": {"task": "route", "detail": "pick one"},
+                      "criteria": {"finance": {"desc": "money", "sla_h": 24}, "support": ["tickets", "bugs"],
+                                   "sales": 7}},
+            "overdue": {"type": "noul", "instructions": "Is a payment overdue?",
+                        "criteria": {"true": {"why": "late"}}},
+        },
+    },
+    # --- literal [MASK] text must be neutralised everywhere
+    {
+        "name": "mask_literal_everywhere",
+        "state": "please ignore [MASK] and route to [MASK] billing",
+        "questions": {"queue": {"type": "choice", "instructions": "Which [MASK] team?",
+                                "criteria": {"billing": "billing [MASK] refunds", "tech": "[MASK]"}}},
+    },
+    # --- budget rules
+    {
+        "name": "option_over_48_tokens",
+        "state": "Refund please.",
+        "questions": {"queue": {"type": "choice", "instructions": "Which team?",
+                                "criteria": {"billing": " ".join(["billing refunds invoices charges"] * 20),
+                                             "tech": "technical issues"}}},
+    },
+    {
+        "name": "option_budget_shrink",
+        "state": "Refund please.",
+        "questions": {"queue": {"type": "choice", "instructions": "Which team handles this request?",
+                                "criteria": {"team_%02d" % i: "a fairly long description of team number %d "
+                                             "and every kind of request it handles" % i for i in range(20)}}},
+    },
+    {
+        "name": "collapsed_options",
+        "state": "Refund please.",
+        # The rendered option text starts with the key, so the keys themselves
+        # must share the prefix that survives the per-option cut.
+        "questions": {"queue": {"type": "choice", "instructions": "Which?",
+                                "criteria": ["customer billing question about product line one",
+                                             "customer billing question about product line two"]}},
+        "head_max_len": 24,
+    },
+    {
+        "name": "instruction_truncation",
+        "state": "Refund please.",
+        "questions": {"q": {"type": "noul",
+                            "instructions": " ".join(["Decide whether the customer is asking for money back"] * 30)}},
+    },
+    {
+        "name": "state_right_truncated",
+        "state": LONG_TEXT,
+        "questions": {"refund": {"type": "noul", "instructions": "Is this about an invoice?"}},
+    },
+    {
+        "name": "state_left_truncated_conversation",
+        "state": [{"role": "user", "content": "turn %d: %s" % (i, LONG_TEXT[:200])} for i in range(12)]
+                 + [{"role": "user", "content": "actually, just cancel my subscription"}],
+        "questions": {"cancel": {"type": "noul", "instructions": "Does the user want to cancel?"}},
+    },
+    {
+        "name": "budget_overrides",
+        "state": LONG_TEXT,
+        "questions": {"queue": {"type": "choice", "instructions": "Which team?",
+                                "criteria": ["finance", "support", "sales"]}},
+        "max_len": 128,
+        "head_max_len": 64,
+    },
+    # --- non-ASCII inside a JSON state
+    {
+        "name": "json_state_non_ascii",
+        "state": {"from": "Zoë", "text": "Die Rechnung ist falsch — bitte korrigieren. 请尽快处理"},
+        "questions": {"queue": {"type": "choice", "instructions": "Which team?",
+                                "criteria": ["billing", "tech", "other"]}},
+    },
+    # --- routing to the multilingual checkpoint
+    {
+        "name": "ml_devanagari",
+        "state": "मुझे इस महीने दो बार शुल्क लिया गया, मुझे पैसे वापस चाहिए",
+        "questions": {"queue": {"type": "choice", "instructions": "Which team?",
+                                "criteria": {"billing": "billing and refunds", "tech": "login and app issues",
+                                             "other": "everything else"}}},
+    },
+    {
+        "name": "ml_khmer",
+        "state": "ខ្ញុំត្រូវបានគិតប្រាក់ពីរដងក្នុងខែនេះ ខ្ញុំចង់បានប្រាក់វិញ",
+        "questions": {"refund": {"type": "noul", "instructions": "Is the customer asking for a refund?"}},
+    },
+    {
+        "name": "ml_cjk",
+        "state": "这个月我被扣了两次钱，我要退款",
+        "questions": {"urgency": {"type": "score", "instructions": "How urgent?",
+                                  "criteria": ["calm", "firm", "angry", "furious"]}},
+    },
+    {
+        "name": "ml_arabic",
+        "state": "تم خصم المبلغ مرتين هذا الشهر، أريد استرداد أموالي",
+        "questions": {"refund": {"type": "noul", "instructions": "Is the customer asking for a refund?"}},
+    },
+    {
+        "name": "ml_latin_non_english",
+        "state": "Me han cobrado dos veces este mes y quiero que me devuelvan el dinero",
+        "questions": {"queue": {"type": "choice", "instructions": "Which team?",
+                                "criteria": ["billing", "tech", "other"]}},
+    },
+    {
+        "name": "ml_mixed_script",
+        "state": "Order #5521 — 注文をキャンセルしたいです. Please help.",
+        "questions": {"cancel": {"type": "noul", "instructions": "Does the user want to cancel?"}},
+    },
+    # --- explicit checkpoint selection through `model`
+    {
+        "name": "model_pinned_multilingual",
+        "state": "I was charged twice this month, I want my money back",
+        "questions": {"refund": {"type": "noul", "instructions": "Is the customer asking for a refund?"}},
+        "model": "multilingual",
+    },
+    {
+        "name": "model_pinned_typed_decisions",
+        "state": "I was charged twice this month, I want my money back",
+        "questions": {"refund": {"type": "noul", "instructions": "Is the customer asking for a refund?"}},
+        "model": "typed-decisions",
+    },
+    {
+        "name": "model_foreign_name_routes_by_default",
+        "state": "I was charged twice this month, I want my money back",
+        "questions": {"refund": {"type": "noul", "instructions": "Is the customer asking for a refund?"}},
+        "model": "jev-1",
+    },
+]
+
+LANG_CASES = [
+    "I was charged twice this month, I want my money back",
+    "Me han cobrado dos veces este mes",
+    "Die Rechnung ist falsch, bitte korrigieren",
+    "Je voudrais annuler ma commande",
+    "मुझे पैसे वापस चाहिए",
+    "ខ្ញុំចង់បានប្រាក់វិញ",
+    "这个月我被扣了两次钱",
+    "注文をキャンセルしたいです",
+    "تم خصم المبلغ مرتين",
+    "Мне дважды списали деньги",
+    "Order #5521 — 注文をキャンセルしたいです. Please help.",
+    "ok",
+    "",
+    "12345 67890",
+    "café résumé naïve",
+    {"subject": "Invoice", "body": "Bitte zahlen Sie bis Freitag"},
+    ["hello", "bonjour", "hola"],
+]
+
+
+class Capture:
+    """Wraps the two Agent internals the fixtures need, per call."""
+
+    def __init__(self):
+        self.items = None
+        self.logits = None
+        self.act = None
+        self.agent = None
+        orig_encode, orig_forward = Agent._encode_state, Agent._forward
+        cap = self
+
+        def encode(agent, *a, **kw):
+            items = orig_encode(agent, *a, **kw)
+            cap.items, cap.agent = items, agent
+            return items
+
+        def forward(agent, b):
+            logits, act = orig_forward(agent, b)
+            cap.logits, cap.act = logits, act
+            return logits, act
+
+        Agent._encode_state, Agent._forward = encode, forward
+
+    def reset(self):
+        self.items = self.logits = self.act = self.agent = None
+
+
+def to_py(x):
+    if isinstance(x, np.ndarray):
+        return x.tolist()
+    if isinstance(x, (np.floating, np.integer)):
+        return x.item()
+    if isinstance(x, dict):
+        return {k: to_py(v) for k, v in x.items()}
+    if isinstance(x, (list, tuple)):
+        return [to_py(v) for v in x]
+    return x
+
+
+def main():
+    ap = argparse.ArgumentParser()
+    ap.add_argument("--out", required=True)
+    args = ap.parse_args()
+    os.makedirs(args.out, exist_ok=True)
+    torch.manual_seed(0)
+    torch.set_num_threads(max(1, (os.cpu_count() or 2) // 2))
+
+    cap = Capture()
+    router = Router(device="cpu", preload=True, max_loaded=3)
+    header = {
+        "laya_version": laya.__version__,
+        "torch_version": torch.__version__,
+        "device": "cpu",
+        "dtype": "float32",
+    }
+
+    cases, checkpoints = [], {}
+    for case in CASES:
+        req = {"state": case["state"], "questions": case["questions"]}
+        kw = {}
+        for key in ("model", "max_len", "head_max_len"):
+            if key in case:
+                req[key] = case[key]
+                if key != "model":
+                    kw[key] = case[key]
+        cap.reset()
+        # The HTTP server, not the Router, maps a foreign model name (a Jev id)
+        # to "let the router choose"; go through the same function.
+        resp = router.predict(case["state"], case["questions"], model=_resolve_model(case.get("model")), **kw)
+        agent = cap.agent
+        rows = []
+        for j, (qid, it) in enumerate(zip(case["questions"], cap.items)):
+            k = len(it["markers"])
+            rows.append({
+                "qid": qid,
+                "input_ids": list(it["ids"]),
+                "markers": list(it["markers"]),
+                "qtype": it["qtype"],
+                "options": it.get("options"),
+                "logits": cap.logits[j, :k].tolist(),
+                "act": cap.act[j].tolist(),
+            })
+        ckpt = resp["routing"]["model"]
+        if ckpt not in checkpoints:
+            checkpoints[ckpt] = {
+                "repo": resp["routing"].get("repo"),
+                "revision": getattr(agent, "revision", None),
+                "cfg": {k: agent.cfg.get(k) for k in ("max_len", "head_max_len", "temperature",
+                                                       "temperature_by_options")},
+                "special_ids": {"cls": agent.tok.cls_token_id, "sep": agent.tok.sep_token_id,
+                                "mask": agent.tok.mask_token_id, "pad": agent.tok.pad_token_id},
+            }
+        cases.append({"name": case["name"], "request": req, "items": rows, "response": to_py(resp)})
+        print("%-40s -> %-16s %d question(s)" % (case["name"], ckpt, len(rows)), file=sys.stderr)
+
+    lang = [{"state": s, "analysis": to_py(analyse(s))} for s in LANG_CASES]
+
+    header["checkpoints"] = checkpoints
+    with open(os.path.join(args.out, "reference.json"), "w") as f:
+        json.dump({"header": header, "cases": cases}, f, ensure_ascii=False, indent=1)
+    with open(os.path.join(args.out, "lang.json"), "w") as f:
+        json.dump({"header": header, "cases": lang}, f, ensure_ascii=False, indent=1)
+
+
+if __name__ == "__main__":
+    main()

script/validate-decision.py

@@ -0,0 +1,133 @@
+#!/usr/bin/env python3
+"""End-to-end check of /v1/systemone decision serving (#341).
+
+Replays every case of the recorded Laya reference corpus
+(`crates/neuron/src/harness/testdata/laya/reference.json`) against a
+serving endpoint -- a neuron directly, a cortex, or the public router --
+and compares each response with what the reference server returned:
+
+  - `routing.model` (which checkpoint answered) and `usage` must match
+    exactly;
+  - each answer's type, and its chosen option or score level, must match,
+    except where the reference itself was within `--tie` of a tie (reduced
+    precision may break a near-tie either way);
+  - probabilities and confidences must be within `--tolerance`. The
+    reference is f32; a bf16 GPU path drifts from it by about as much as
+    the reference's own bf16 GPU run does.
+
+It prints each case's latency and a summary, and exits non-zero on any
+mismatch. Nothing is loaded or unloaded: point it at an endpoint that
+already serves a decision model.
+
+Usage:
+  script/validate-decision.py URL [--model M] [--bearer KEY]
+
+  URL      http://quadbrat.hanzalova.internal:13131   (neuron)
+           http://hanzalova.internal:31313             (cortex)
+           https://router.example                      (public chain)
+  --model  override every request's `model` (e.g. helexa/one); by
+           default each case sends what it was recorded with, which
+           includes checkpoint pins and a foreign Jev model name.
+"""
+import argparse
+import json
+import os
+import statistics
+import sys
+import time
+import urllib.error
+import urllib.request
+
+FIXTURE = os.path.join(os.path.dirname(os.path.abspath(__file__)), "..", "crates", "neuron", "src",
+                       "harness", "testdata", "laya", "reference.json")
+
+
+def post(url, body, bearer):
+    req = urllib.request.Request(url, data=json.dumps(body).encode(), method="POST",
+                                 headers={"Content-Type": "application/json"})
+    if bearer:
+        req.add_header("Authorization", "Bearer " + bearer)
+    started = time.perf_counter()
+    try:
+        with urllib.request.urlopen(req, timeout=120) as r:
+            payload = json.load(r)
+            status = r.status
+    except urllib.error.HTTPError as e:
+        payload = json.loads(e.read() or b"{}")
+        status = e.code
+    return status, payload, (time.perf_counter() - started) * 1e3
+
+
+def top_gap(probs):
+    p = sorted(probs, reverse=True)
+    return p[0] - p[1] if len(p) > 1 else 1.0
+
+
+def compare(name, got, want, tol, tie):
+    errs = []
+    if got.get("routing", {}).get("model") != want["routing"]["model"]:
+        errs.append("routing.model %r != %r" % (got.get("routing", {}).get("model"), want["routing"]["model"]))
+    if got.get("usage") != want["usage"]:
+        errs.append("usage %r != %r" % (got.get("usage"), want["usage"]))
+    ga, wa = got.get("answers", {}), want["answers"]
+    if list(ga) != list(wa):
+        errs.append("answer ids %r != %r" % (list(ga), list(wa)))
+        return errs, 0.0
+    worst = 0.0
+    for qid, w in wa.items():
+        g = ga[qid]
+        if g.get("type") != w["type"]:
+            errs.append("%s: type %r != %r" % (qid, g.get("type"), w["type"]))
+            continue
+        wp = w.get("probabilities")
+        near_tie = wp is not None and top_gap(list(wp.values())) < tie
+        if w["type"] == "choice" and g.get("choice") != w["choice"] and not near_tie:
+            errs.append("%s: choice %r != %r" % (qid, g.get("choice"), w["choice"]))
+        if wp:
+            for opt, p in wp.items():
+                worst = max(worst, abs(g["probabilities"][opt] - p))
+        for key in ("noul", "confidence", "answer_confidence"):
+            if key in w:
+                worst = max(worst, abs(g[key] - w[key]))
+    if worst > tol:
+        errs.append("max |Δ| %.4f > %.4f" % (worst, tol))
+    return errs, worst
+
+
+def main():
+    ap = argparse.ArgumentParser()
+    ap.add_argument("url")
+    ap.add_argument("--model")
+    ap.add_argument("--bearer")
+    ap.add_argument("--tolerance", type=float, default=0.1)
+    ap.add_argument("--tie", type=float, default=0.05)
+    args = ap.parse_args()
+    endpoint = args.url.rstrip("/") + "/v1/systemone"
+    cases = json.load(open(FIXTURE))["cases"]
+
+    failures, latencies, worst_all = 0, [], 0.0
+    for case in cases:
+        body = dict(case["request"])
+        if args.model:
+            body["model"] = args.model
+        status, got, ms = post(endpoint, body, args.bearer)
+        latencies.append(ms)
+        if status != 200:
+            failures += 1
+            print("FAIL %-40s HTTP %d %s" % (case["name"], status, json.dumps(got)[:200]))
+            continue
+        errs, worst = compare(case["name"], got, case["response"], args.tolerance, args.tie)
+        worst_all = max(worst_all, worst)
+        mark = "ok  " if not errs else "FAIL"
+        failures += bool(errs)
+        print("%s %-40s %7.1f ms  %-15s max|Δ| %.4f  %s" % (
+            mark, case["name"], ms, got.get("routing", {}).get("model", "?"), worst, "; ".join(errs)))
+
+    print("\n%d/%d cases ok; latency median %.1f ms, p90 %.1f ms; worst |Δ| %.4f" % (
+        len(cases) - failures, len(cases), statistics.median(latencies),
+        sorted(latencies)[int(0.9 * (len(latencies) - 1))], worst_all))
+    sys.exit(1 if failures else 0)
+
+
+if __name__ == "__main__":
+    main()

script/validate-image.sh

@@ -9,17 +9,27 @@
 # image modality.
 #
 # Usage:
-#   script/validate-image.sh [host] [model_id]
+#   script/validate-image.sh [host] [model_id] [quant]
 #
 # Defaults:
 #   host     = benjy.hanzalova.internal   (the v1 image placement)
 #   model_id = Tongyi-MAI/Z-Image-Turbo
+#   quant    = none (BF16). A 12 GB card needs q8_0.
+#
+# A co-resident model evicted for the probe window is restored on every
+# exit path, failures included.
 
 set -euo pipefail
 
 HOST="${1:-benjy.hanzalova.internal}"
 MODEL_ID="${2:-Tongyi-MAI/Z-Image-Turbo}"
+QUANT="${3:-}"
 BASE="http://${HOST}:13131"
+if [[ -n "${QUANT}" ]]; then
+  LOAD_JSON="{\"model_id\": \"${MODEL_ID}\", \"harness\": \"candle\", \"quant\": \"${QUANT}\"}"
+else
+  LOAD_JSON="{\"model_id\": \"${MODEL_ID}\", \"harness\": \"candle\"}"
+fi
 
 echo "==> /version"
 curl -sf -m 10 "${BASE}/version" | python3 -c 'import json,sys; b=json.load(sys.stdin); print("  ", b.get("git_sha","?")[:12], b.get("profile"), b.get("features"))'
@@ -28,26 +38,47 @@ curl -sf -m 10 "${BASE}/version" | python3 -c 'import json,sys; b=json.load(sys.
 # cortex's job. On a host whose resident text model leaves too little
 # VRAM for the DiT (benjy: 8B @ 16 GB + 14.6 GB DiT > 24 GB), evict
 # it for the probe window and restore it afterwards.
-RESTORE_JSON=""
-EVICT=$(curl -sf -m 10 "${BASE}/models" | python3 -c '
+EVICTED=()
+# Unload the probe model and restore the evicted ones however the script
+# ends: a failed generation must not leave the host without its models.
+restore() {
+  local rc=$?
+  curl -s -m 60 -X POST "${BASE}/models/unload" \
+    -H 'Content-Type: application/json' \
+    -d "{\"model_id\": \"${MODEL_ID}\"}" > /dev/null || true
+  for id in "${EVICTED[@]}"; do
+    echo "==> restoring ${id}"
+    curl -s -m 900 -X POST "${BASE}/models/load" \
+      -H 'Content-Type: application/json' \
+      -d "{\"model_id\": \"${id}\", \"harness\": \"candle\"}" \
+      || echo "   RESTORE FAILED: reload ${id} by hand"
+    echo
+  done
+  exit "${rc}"
+}
+trap restore EXIT
+
+mapfile -t RESIDENT < <(curl -sf -m 10 "${BASE}/models" | python3 -c '
 import json, sys
-models = json.load(sys.stdin)
-loaded = [m for m in models if m["status"] == "loaded" and "image" not in m.get("capabilities", [])]
-print(loaded[0]["id"] if loaded else "")
+for m in json.load(sys.stdin):
+    if m["status"] == "loaded" and "image" not in m.get("capabilities", []):
+        print(m["id"])
 ')
-if [[ -n "${EVICT}" ]]; then
-  echo "==> evicting co-resident ${EVICT} for the probe window"
+for id in "${RESIDENT[@]}"; do
+  [[ -z "${id}" ]] && continue
+  echo "==> evicting co-resident ${id} for the probe window"
   curl -sf -m 120 -X POST "${BASE}/models/unload" \
     -H 'Content-Type: application/json' \
-    -d "{\"model_id\": \"${EVICT}\"}"
+    -d "{\"model_id\": \"${id}\"}"
   echo
-  RESTORE_JSON="{\"model_id\": \"${EVICT}\", \"harness\": \"candle\"}"
-fi
+  EVICTED+=("${id}")
+done
+
 
 echo "==> load ${MODEL_ID}"
 curl -sf -m 900 -X POST "${BASE}/models/load" \
   -H 'Content-Type: application/json' \
-  -d "{\"model_id\": \"${MODEL_ID}\", \"harness\": \"candle\"}"
+  -d "${LOAD_JSON}"
 echo
 
 echo "==> capabilities"
@@ -80,20 +111,6 @@ t = usage["helexa_timing"]
 print("   512x512 ok | units={:.3f} | encode={}ms denoise={}ms decode={}ms".format(units, t["encode_ms"], t["denoise_ms"], t["decode_ms"]))
 '
 
-echo "==> unload"
-curl -sf -m 60 -X POST "${BASE}/models/unload" \
-  -H 'Content-Type: application/json' \
-  -d "{\"model_id\": \"${MODEL_ID}\"}"
-echo
-
-if [[ -n "${RESTORE_JSON}" ]]; then
-  echo "==> restoring ${EVICT}"
-  curl -sf -m 900 -X POST "${BASE}/models/load" \
-    -H 'Content-Type: application/json' \
-    -d "${RESTORE_JSON}"
-  echo
-fi
-
 echo "==> health"
 curl -sf -m 10 "${BASE}/health" > /dev/null && echo "   healthy"
-echo "PASS"
+echo "PASS (unload and restore follow)"