diff --git a/crates/ml/src/cuda_pipeline/gpu_dqn_trainer.rs b/crates/ml/src/cuda_pipeline/gpu_dqn_trainer.rs
index 6da91486c..3ad4d299a 100644
--- a/crates/ml/src/cuda_pipeline/gpu_dqn_trainer.rs
+++ b/crates/ml/src/cuda_pipeline/gpu_dqn_trainer.rs
@@ -2143,11 +2143,30 @@ impl GpuDqnTrainer {
         };
         let grad_norm = self.debug_buffer_norm_f32(self.ptrs.grad_buf, self.total_params).unwrap_or(f32::NAN);
 
+        // Raw DtoH probe: read first 4 floats directly from grad_buf and grad_norm_buf.
+        // Bypasses all kernel infrastructure to verify the buffer actually has data.
+        let mut raw_grad = [0.0_f32; 4];
+        let mut raw_norm = [0.0_f32; 1];
+        unsafe {
+            cudarc::driver::sys::cuStreamSynchronize(self.stream.cu_stream());
+            cudarc::driver::sys::cuMemcpyDtoH_v2(
+                raw_grad.as_mut_ptr().cast(), self.ptrs.grad_buf, 16,
+            );
+            cudarc::driver::sys::cuMemcpyDtoH_v2(
+                raw_norm.as_mut_ptr().cast(), self.ptrs.grad_norm_buf, 4,
+            );
+        }
+
         tracing::warn!(
             step,
             d_val_norm,
             d_adv_norm,
             grad_norm,
+            raw_grad_0 = raw_grad[0],
+            raw_grad_1 = raw_grad[1],
+            raw_grad_2 = raw_grad[2],
+            raw_grad_3 = raw_grad[3],
+            raw_norm = raw_norm[0],
             c51_alpha = self.c51_alpha,
             "BUFFER_DIAG: per-step buffer norms after forward+backward"
         );