diff --git a/crates/ml/src/cuda_pipeline/experience_kernels.cu b/crates/ml/src/cuda_pipeline/experience_kernels.cu
index b74d69f95..42051e5bb 100644
--- a/crates/ml/src/cuda_pipeline/experience_kernels.cu
+++ b/crates/ml/src/cuda_pipeline/experience_kernels.cu
@@ -1010,13 +1010,22 @@ extern "C" __global__ void experience_action_select(
 
     /* Output Q-gap for conviction-based position sizing.
      * Q-gap = Q(best_direction) - Q(Flat). Higher = more conviction.
-     * Used by env_step to scale target_position. */
+     * Used by env_step to scale target_position.
+     *
+     * D7/N7: During contrarian override, the action selected is argmin rather
+     * than argmax. The raw Q(best) - Q(Flat) gap would mislead the position
+     * sizer into full conviction for what's actually the "deliberately wrong"
+     * action. Zero out the gap so contrarian trades size minimal. */
     if (out_q_gaps != NULL) {
-        int flat_idx = 3; /* Flat = index 3 for direction(4): Short(0)/Hold(1)/Long(2)/Flat(3) */
-        float q_best = q_b0[argmax_n(q_b0, b0_size)];
-        float q_flat = q_b0[flat_idx];
-        float gap = q_best - q_flat;
-        out_q_gaps[i] = (gap > 0.0f) ? gap : 0.0f;
+        if (contrarian_active != 0) {
+            out_q_gaps[i] = 0.0f;
+        } else {
+            int flat_idx = 3; /* Flat = index 3 for direction(4): Short(0)/Hold(1)/Long(2)/Flat(3) */
+            float q_best = q_b0[argmax_n(q_b0, b0_size)];
+            float q_flat = q_b0[flat_idx];
+            float gap = q_best - q_flat;
+            out_q_gaps[i] = (gap > 0.0f) ? gap : 0.0f;
+        }
     }
 }