diff --git a/crates/ml-alpha/src/trainer/integrated.rs b/crates/ml-alpha/src/trainer/integrated.rs
index 45f1ff9ac..b541062f4 100644
--- a/crates/ml-alpha/src/trainer/integrated.rs
+++ b/crates/ml-alpha/src/trainer/integrated.rs
@@ -5445,11 +5445,20 @@ impl IntegratedTrainer {
             }
         }
 
-        // ── Confidence gate — override opening actions to Hold when C51
-        // distributional Q has high uncertainty for the chosen action.
-        // Fires AFTER π samples an action but BEFORE trail/heat/market
-        // pipeline. Reads pos_state from lobsim (pre-snapshot = current
-        // position) to only gate on flat entries.
+            if capturing_prefill {
+                let graph = self
+                    .stream
+                    .end_capture(
+                        CUgraphInstantiate_flags::CUDA_GRAPH_INSTANTIATE_FLAG_AUTO_FREE_ON_LAUNCH,
+                    )
+                    .context("prefill end_capture")?
+                    .ok_or_else(|| anyhow::anyhow!("prefill end_capture returned None"))?;
+                self.prefill_graph = Some(graph);
+            }
+        } // end else (warmup / capture dispatch)
+
+        // ── Gates + log_pi: OUTSIDE graph capture so they fire every
+        // step (not captured as no-ops during warmup). Device-side kernels.
         {
             let pos_d_ref: &CudaSlice<u8> = lobsim.pos_d();
             let b_size_i = b_size as i32;
@@ -5472,8 +5481,6 @@ impl IntegratedTrainer {
                 ).map_err(|e| anyhow::anyhow!("rl_confidence_gate: {:?}", e))?;
             }
         }
-
-        // FRD gate — same warmup as confidence gate.
         {
             let pos_d_ref: &CudaSlice<u8> = lobsim.pos_d();
             let b_size_i = b_size as i32;
@@ -5495,11 +5502,6 @@ impl IntegratedTrainer {
                 ).map_err(|e| anyhow::anyhow!("rl_frd_gate: {:?}", e))?;
             }
         }
-
-        // log pi_old at the (possibly gate-overridden) action — AFTER
-        // confidence gate + FRD gate so that log π_old matches the action
-        // that actually executes, giving PPO a correct importance ratio
-        // when a gate overrides the sampled action to Hold.
         {
             let grid_x = ((b_size as u32) + 31) / 32;
             let b_size_i = b_size as i32;
@@ -5515,22 +5517,10 @@ impl IntegratedTrainer {
                     (grid_x.max(1), 1, 1), (32, 1, 1), 0,
                     self.raw_stream,
                     &mut ptrs[..args.len()],
-                ).map_err(|e| anyhow::anyhow!("log_pi_at_action: {:?}", e))?;
+                ).map_err(|e| anyhow::anyhow!("log_pi_at_action (post-gate): {:?}", e))?;
             }
         }
 
-            if capturing_prefill {
-                let graph = self
-                    .stream
-                    .end_capture(
-                        CUgraphInstantiate_flags::CUDA_GRAPH_INSTANTIATE_FLAG_AUTO_FREE_ON_LAUNCH,
-                    )
-                    .context("prefill end_capture")?
-                    .ok_or_else(|| anyhow::anyhow!("prefill end_capture returned None"))?;
-                self.prefill_graph = Some(graph);
-            }
-        } // end else (warmup / capture dispatch)
-
         // ── Step 5 (Phase R6): GPU-pure env step. ─────────────────────
         // No per-batch host loop — actions land in lobsim's
         // market_targets_d via the `actions_to_market_targets` kernel
@@ -7077,7 +7067,19 @@ impl IntegratedTrainer {
             }
         }
 
-        // Confidence gate
+            if capturing_prefill {
+                let graph = self
+                    .stream
+                    .end_capture(
+                        CUgraphInstantiate_flags::CUDA_GRAPH_INSTANTIATE_FLAG_AUTO_FREE_ON_LAUNCH,
+                    )
+                    .context("prefill end_capture")?
+                    .ok_or_else(|| anyhow::anyhow!("prefill end_capture returned None"))?;
+                self.prefill_graph = Some(graph);
+            }
+        } // end else (warmup / capture dispatch)
+
+        // ── Gates + log_pi: OUTSIDE graph capture (GPU data path).
         {
             let pos_d_ref: &CudaSlice<u8> = lobsim.pos_d();
             let b_size_i = b_size as i32;
@@ -7097,11 +7099,9 @@ impl IntegratedTrainer {
                     (b_size as u32, 1, 1), (1, 1, 1), 0,
                     self.raw_stream,
                     &mut ptrs[..args.len()],
-                ).map_err(|e| anyhow::anyhow!("rl_confidence_gate: {:?}", e))?;
+                ).map_err(|e| anyhow::anyhow!("rl_confidence_gate (gpu): {:?}", e))?;
             }
         }
-
-        // FRD gate
         {
             let pos_d_ref: &CudaSlice<u8> = lobsim.pos_d();
             let b_size_i = b_size as i32;
@@ -7120,14 +7120,9 @@ impl IntegratedTrainer {
                     (b_size as u32, 1, 1), (1, 1, 1), 0,
                     self.raw_stream,
                     &mut ptrs[..args.len()],
-                ).map_err(|e| anyhow::anyhow!("rl_frd_gate: {:?}", e))?;
+                ).map_err(|e| anyhow::anyhow!("rl_frd_gate (gpu): {:?}", e))?;
             }
         }
-
-        // log pi_old at the (possibly gate-overridden) action — AFTER
-        // confidence gate + FRD gate so that log π_old matches the action
-        // that actually executes, giving PPO a correct importance ratio
-        // when a gate overrides the sampled action to Hold.
         {
             let grid_x = ((b_size as u32) + 31) / 32;
             let b_size_i = b_size as i32;
@@ -7143,22 +7138,10 @@ impl IntegratedTrainer {
                     (grid_x.max(1), 1, 1), (32, 1, 1), 0,
                     self.raw_stream,
                     &mut ptrs[..args.len()],
-                ).map_err(|e| anyhow::anyhow!("log_pi_at_action: {:?}", e))?;
+                ).map_err(|e| anyhow::anyhow!("log_pi_at_action (post-gate gpu): {:?}", e))?;
             }
         }
 
-            if capturing_prefill {
-                let graph = self
-                    .stream
-                    .end_capture(
-                        CUgraphInstantiate_flags::CUDA_GRAPH_INSTANTIATE_FLAG_AUTO_FREE_ON_LAUNCH,
-                    )
-                    .context("prefill end_capture")?
-                    .ok_or_else(|| anyhow::anyhow!("prefill end_capture returned None"))?;
-                self.prefill_graph = Some(graph);
-            }
-        } // end else (warmup / capture dispatch)
-
         // ── Step 5: GPU-pure env step (GPU data path). ─────────────────
         //
         // apply_snapshot from device: copy batch 0's last-snapshot book