diff --git a/crates/ml/src/cuda_pipeline/gpu_backtest_evaluator.rs b/crates/ml/src/cuda_pipeline/gpu_backtest_evaluator.rs
index 68bda2efa..0c561fcd8 100644
--- a/crates/ml/src/cuda_pipeline/gpu_backtest_evaluator.rs
+++ b/crates/ml/src/cuda_pipeline/gpu_backtest_evaluator.rs
@@ -1205,7 +1205,11 @@ impl GpuBacktestEvaluator {
             }
 
             // ── Phase 4: Greedy action selection on chunked Q-values ─────────
-            // Kernel signature: (q_values, out_actions, rng_states, out_q_gaps, epsilon, N, b0, b1, b2, q_gap_threshold)
+            // Kernel signature: (q_values, out_actions, rng_states, out_q_gaps, epsilon, N,
+            //                    b0, b1, b2, q_gap_threshold, portfolio_states, min_hold_bars, max_position)
+            let null_portfolio: u64 = 0;  // NULL — backtest has no hold enforcement yet
+            let eval_min_hold: i32 = 0;   // 0 = disabled during evaluation
+            let eval_max_pos: f32 = 0.0;
             unsafe {
                 self.stream
                     .launch_builder(as_kernel)
@@ -1219,6 +1223,9 @@ impl GpuBacktestEvaluator {
                     .arg(&b1)
                     .arg(&b2)
                     .arg(&self.q_gap_threshold)
+                    .arg(&null_portfolio)       // portfolio_states (NULL — no hold masking in eval)
+                    .arg(&eval_min_hold)        // min_hold_bars (0 = disabled)
+                    .arg(&eval_max_pos)         // max_position (unused when disabled)
                     .launch(launch_cfg)
                     .map_err(|e| MLError::ModelError(format!(
                         "chunked action_select chunk_start={chunk_start}: {e}"