diff --git a/crates/ml/tests/gpu_backtest_validation.rs b/crates/ml/tests/gpu_backtest_validation.rs
index ebdef2932..2853d2d0b 100644
--- a/crates/ml/tests/gpu_backtest_validation.rs
+++ b/crates/ml/tests/gpu_backtest_validation.rs
@@ -284,7 +284,14 @@ mod gpu_tests {
         // env_step decoder. Match production eval-baseline.rs.
         evaluator.set_branch_sizes(&DqnBacktestConfig::from_network_dims((256, 256, 128, 128)));
 
-        let model = constant_action_model(4, &stream); // Always Long100
+        // Action 72 = Long×Full×order=0×urgency=0 under 4-3-3-3 factored
+        // decoding (action = dir*27 + mag*9 + order*3 + urgency = 2*27 +
+        // 2*9 + 0 + 0). Pre-Phase-8.5 the eval used a flat 4-action enum
+        // where `4` reportedly meant Long100, but SP21 Phase 8.5
+        // (2026-05-12) wired the factored decoder where `4` now decodes
+        // to Short-Quarter (dir=0, mag=0). The "Always Long100" comment
+        // was correct intent but the constant wasn't migrated.
+        let model = constant_action_model(72, &stream); // Always Long100 (factored)
         let metrics = evaluator
             .evaluate(&model, 24)
             .expect("evaluation should succeed");
@@ -405,7 +412,9 @@ mod gpu_tests {
         // env_step decoder. Match production eval-baseline.rs.
         evaluator.set_branch_sizes(&DqnBacktestConfig::from_network_dims((256, 256, 128, 128)));
 
-        let model = constant_action_model(4, &stream); // Always Long100
+        // Action 72 = Long×Full×order=0×urgency=0 under 4-3-3-3 factored
+        // decoding (see test_always_long_on_downtrend for full derivation).
+        let model = constant_action_model(72, &stream); // Always Long100 (factored)
         let metrics = evaluator
             .evaluate(&model, 24)
             .expect("evaluation should succeed");