diff --git a/crates/ml/src/cuda_pipeline/gpu_experience_collector.rs b/crates/ml/src/cuda_pipeline/gpu_experience_collector.rs
index 956f50f02..89af4227b 100644
--- a/crates/ml/src/cuda_pipeline/gpu_experience_collector.rs
+++ b/crates/ml/src/cuda_pipeline/gpu_experience_collector.rs
@@ -2051,20 +2051,28 @@ impl GpuExperienceCollector {
                     )))?;
             }
 
-            // One-shot diagnostic: after first state_gather, readback batch_states[0] to verify OFI
+            // One-shot diagnostic: readback batch_states[0..state_dim] via pinned memory
             if t == 0 && self.last_experience_count == 0 {
                 let sd = self.state_dim;
-                let mut state0 = vec![0.0_f32; sd];
-                if let Ok(()) = self.stream.memcpy_dtoh(&self.batch_states, &mut state0) {
-                    let ofi_slice = &state0[66..84.min(sd)];
-                    let ofi_nonzero = ofi_slice.iter().filter(|&&v| v != 0.0).count();
-                    info!(
-                        "STATE_GATHER_DIAG: state[66..84]={:?}, nonzero={}/{}, ofi_ptr=0x{:x}, ofi_dim={}, total_bars={}",
-                        ofi_slice, ofi_nonzero, ofi_slice.len(),
-                        self.ofi_gpu.device_ptr(&self.stream).0,
-                        self.ofi_dim, config.total_bars,
+                let pinned = unsafe { PinnedHostBuf::<f32>::new(sd)
+                    .map_err(|e| MLError::ModelError(format!("diag pinned alloc: {e}")))? };
+                unsafe {
+                    cudarc::driver::sys::cuMemcpyDtoHAsync_v2(
+                        pinned.ptr as *mut std::ffi::c_void,
+                        self.batch_states.device_ptr(&self.stream).0,
+                        (sd * 4) as usize,
+                        self.stream.cu_stream(),
                     );
+                    cudarc::driver::sys::cuStreamSynchronize(self.stream.cu_stream());
                 }
+                let state0 = unsafe { std::slice::from_raw_parts(pinned.ptr, sd) };
+                let ofi_slice = &state0[66..84.min(sd)];
+                let ofi_nonzero = ofi_slice.iter().filter(|&&v| v != 0.0).count();
+                info!(
+                    "STATE_GATHER_DIAG: state[66..84]={:?}, nonzero={}/{}, ofi_dim={}, total_bars={}",
+                    ofi_slice, ofi_nonzero, ofi_slice.len(),
+                    self.ofi_dim, config.total_bars,
+                );
             }
 
             // ── 2. cuBLAS Q-forward: batch_states → logits ──────────────