diff --git a/Cargo.lock b/Cargo.lock
index aa6200a..e618f9d 100644
--- a/Cargo.lock
+++ b/Cargo.lock
@@ -472,6 +472,7 @@ dependencies = [
  "anyhow",
  "cust",
  "engine-cpu",
+ "hex",
  "log",
  "metrics",
  "num-bigint",
diff --git a/crates/engine-gpu-cuda/Cargo.toml b/crates/engine-gpu-cuda/Cargo.toml
index 20102da..48b2fc1 100644
--- a/crates/engine-gpu-cuda/Cargo.toml
+++ b/crates/engine-gpu-cuda/Cargo.toml
@@ -23,6 +23,7 @@ primitive-types = { workspace = true }
 log = { workspace = true }
 thiserror = { workspace = true }
 anyhow = { workspace = true }
+hex = { workspace = true }
 # Metrics (optional; used for emitting generic counters/gauges from this engine)
 metrics = { path = "../metrics", optional = true }
 
diff --git a/crates/engine-gpu-cuda/src/kernels/qpow_kernel.cu b/crates/engine-gpu-cuda/src/kernels/qpow_kernel.cu
index 5a97c93..c9ee6e2 100644
--- a/crates/engine-gpu-cuda/src/kernels/qpow_kernel.cu
+++ b/crates/engine-gpu-cuda/src/kernels/qpow_kernel.cu
@@ -538,6 +538,9 @@ extern "C" __global__ void qpow_montgomery_g2_kernel(
     int*            __restrict__ found_flag,    // 0 -> not found, 1 -> found
     uint32_t*       __restrict__ out_index,     // linear index (t * iters + j)
     uint8_t*        __restrict__ out_distance_be, // 64 bytes
+    // Debug output buffers (optional; host may pass nullptrs)
+    uint8_t*        __restrict__ out_dbg_y_be,    // 64 bytes (optional)
+    uint8_t*        __restrict__ out_dbg_h_be,    // 64 bytes (optional)
  
     // Threading parameters
     const uint32_t num_threads,
@@ -648,15 +651,27 @@ extern "C" __global__ void qpow_montgomery_g2_kernel(
                 if (out_index) {
                     *out_index = tid * iters + j;
                 }
-                // Write distance
+                // Write distance and debug buffers (if provided)
                 if (out_distance_be) {
-uint8_t dist_out[64];
+                    uint8_t dist_out[64];
                     be_u64x8_to_be64_bytes(dist_limb, dist_out);
 #pragma unroll
                     for (int i = 0; i < 64; ++i) {
                         out_distance_be[i] = dist_out[i];
                     }
                 }
+                if (out_dbg_y_be) {
+#pragma unroll
+                    for (int i = 0; i < 64; ++i) {
+                        out_dbg_y_be[i] = y_be[i];
+                    }
+                }
+                if (out_dbg_h_be) {
+#pragma unroll
+                    for (int i = 0; i < 64; ++i) {
+                        out_dbg_h_be[i] = h_be[i];
+                    }
+                }
             }
             return; // early-exit after claiming
         }
diff --git a/crates/engine-gpu-cuda/src/lib.rs b/crates/engine-gpu-cuda/src/lib.rs
index 03d9588..5362c9c 100644
--- a/crates/engine-gpu-cuda/src/lib.rs
+++ b/crates/engine-gpu-cuda/src/lib.rs
@@ -358,6 +358,10 @@ impl CudaEngine {
                     .with_context(|| "alloc/copy d_index")?;
                 let d_distance = cuda::memory::DeviceBuffer::<u8>::zeroed(64)
                     .with_context(|| "alloc d_distance")?;
+                let d_dbg_y = cuda::memory::DeviceBuffer::<u8>::zeroed(64)
+                    .with_context(|| "alloc d_dbg_y")?;
+                let d_dbg_h = cuda::memory::DeviceBuffer::<u8>::zeroed(64)
+                    .with_context(|| "alloc d_dbg_h")?;
 
                 // Launch G2 kernel
                 // Attempt to populate __constant__ symbols for per-job constants (G2 fast path).
@@ -475,6 +479,8 @@ impl CudaEngine {
                         d_found.as_device_ptr(),
                         d_index.as_device_ptr(),
                         d_distance.as_device_ptr(),
+                        d_dbg_y.as_device_ptr(),
+                        d_dbg_h.as_device_ptr(),
                         num_threads as u32,
                         iters_per_thread as u32
                     ))
@@ -490,15 +496,7 @@ impl CudaEngine {
                     .map(|v| v == "1" || v.eq_ignore_ascii_case("true"))
                     .unwrap_or(false)
                 {
-                    if let (
-                        Ok(mut c_en),
-                        Ok(mut c_idx),
-                        Ok(mut c_dec),
-                        Ok(mut c_y),
-                        Ok(mut c_h),
-                        Ok(mut c_t),
-                        Ok(mut c_th),
-                    ) = (
+                    if let (Ok(c_en), Ok(c_idx), Ok(c_dec), Ok(c_y), Ok(c_h), Ok(c_t), Ok(c_th)) = (
                         module.get_global::<i32>(CString::new("C_SAMPLER_ENABLE")?.as_c_str()),
                         module.get_global::<u32>(CString::new("C_SAMPLER_INDEX")?.as_c_str()),
                         module.get_global::<u32>(CString::new("C_SAMPLER_DECISION")?.as_c_str()),
@@ -589,10 +587,24 @@ impl CudaEngine {
                         });
                     } else {
                         // Treat as not found: account coverage and advance like the "not found" path
-                        log::warn!(target: "miner", "CUDA G2: candidate rejected by host re-verification (false positive): idx={k}");
+                        // Copy debug y/h bytes from device and log first 16 bytes (hex) for diagnosis
+                        let mut dbg_y = [0u8; 64];
+                        let mut dbg_h = [0u8; 64];
+                        // Ignore copy errors in logging path; continue accounting regardless
+                        let _ = d_dbg_y.copy_to(&mut dbg_y);
+                        let _ = d_dbg_h.copy_to(&mut dbg_h);
+                        let y_hex = hex::encode(&dbg_y[..16]);
+                        let h_hex = hex::encode(&dbg_h[..16]);
+                        log::warn!(
+                            target: "miner",
+                            "CUDA G2: candidate rejected by host re-verification (false positive): idx={k}, y[0..16]={}, h[0..16]={}",
+                            y_hex,
+                            h_hex
+                        );
                         #[cfg(feature = "metrics")]
                         {
                             metrics::inc_candidates_false_positive("gpu-cuda");
+                            metrics::inc_sample_mismatch("gpu-cuda");
                         }
                         hash_count = hash_count.saturating_add(covered);
                         if covered < total_elems_u64 {