diff --git a/crates/ml/src/cuda_pipeline/gpu_dqn_trainer.rs b/crates/ml/src/cuda_pipeline/gpu_dqn_trainer.rs
index 55a713ac3..44aa51971 100644
--- a/crates/ml/src/cuda_pipeline/gpu_dqn_trainer.rs
+++ b/crates/ml/src/cuda_pipeline/gpu_dqn_trainer.rs
@@ -8219,9 +8219,13 @@ impl GpuDqnTrainer {
                 12 * std::mem::size_of::<f32>(), self.stream.cu_stream(),
             ).map_err(|e| MLError::ModelError(format!("q_out sample0 DtoD: {e}")))?;
         }
-        // No sync — CPU reads previous step's values from pinned host pointer.
+        // Sync: populate_q_out + q_stats_reduce just launched — must wait for
+        // pinned memory writes to complete before CPU reads.
+        unsafe {
+            let rc = cudarc::driver::sys::cuStreamSynchronize(self.stream.cu_stream());
+            assert_eq!(rc, cudarc::driver::sys::cudaError_enum::CUDA_SUCCESS, "q_stats stream sync");
+        }
 
-        // Q-mean EMA update — AFTER stream sync so pinned memory has the GPU-written value.
         self.update_q_mean_ema();
 
         // Read from pinned buffer — host[0..7] = q_stats, host[7..19] = q_out sample 0