diff --git a/crates/ml/src/cuda_pipeline/trade_physics.cuh b/crates/ml/src/cuda_pipeline/trade_physics.cuh
index 5fe68bb90..f72d98785 100644
--- a/crates/ml/src/cuda_pipeline/trade_physics.cuh
+++ b/crates/ml/src/cuda_pipeline/trade_physics.cuh
@@ -237,13 +237,16 @@ __device__ __forceinline__ float compute_drawdown(
  * runs where no trades have completed yet — without priors Kelly would
  * be undefined for the first ~10 trades.
  *
- * The safety_multiplier is expected to be health-coupled by the caller:
- *   safety = 0.5 + 0.5 × health
- * At health=1 (healthy policy): full Kelly — trust the learned policy.
- * At health=0 (collapsing):     half Kelly — constrain when decisions are
- * less reliable. Same temporal-coupling pattern as the distillation SAXPY
- * alpha and the Q-target label smoothing eps. Applied uniformly in both
- * training and backtest envs so there's no train/val mismatch.
+ * The safety_multiplier is composed by the caller from two orthogonal
+ * adaptive signals:
+ *   safety = max(0.5 + 0.5 × health,  conviction)
+ *                 \---------------/  \---------/
+ *                 training-stability  per-sample
+ *                 floor               policy confidence
+ * Bounded to [0.5, 1.0] by the health-safety floor. Same temporal-coupling
+ * pattern as the distillation SAXPY alpha and the Q-target label smoothing
+ * eps. Applied uniformly in both training and backtest envs so there's no
+ * train/val mismatch.
  *
  * The `conviction` parameter (∈ [0, 1], caller-normalised) drives the
  * cold-start warmup floor — no longer a static 0.5f. Derived by the
@@ -608,12 +611,36 @@ __device__ __forceinline__ void unified_env_step_core(
         );
     }
 
-    /* ── Step 5: Kelly cap (health-coupled safety + conviction-adaptive
-     * warmup floor; uses physics scale). Conviction flows per-sample from
-     * the action_select kernel's direction Q-spread. */
+    /* ── Step 5: Kelly cap (safety composed from two orthogonal signals
+     * + conviction-adaptive warmup floor; uses physics scale).
+     *
+     * safety_multiplier = max(health_safety, conviction)
+     *   health_safety    = 0.5 + 0.5 × health     [training-stability floor]
+     *   conviction       ∈ [0, 1]                 [per-sample policy confidence]
+     *
+     * Health measures training stability (is the optimiser making coherent
+     * progress?). Conviction measures per-state policy confidence (is the
+     * Q-spread clearly favouring the taken direction over pathology bins?).
+     * These are orthogonal: a policy can be confident on a given state even
+     * before training globally stabilises, and a stable training regime can
+     * still produce low-conviction per-state decisions. max() composes them
+     * conservatively — the cap uses whichever signal says "trust the policy
+     * more" at this sample. Bounded to [0.5, 1.0] by the health floor.
+     *
+     * Conviction flows per-sample from the action_select kernel's direction
+     * Q-spread. Both signals are adaptive / temporal. No static knobs.
+     *
+     * Per project_magnitude_eval_collapse_kelly_capped.md: at the smoke-test
+     * horizon, typical health=0.49 gives health_safety=0.745 which sits
+     * coincidentally on the Half/Full decoder boundary (abs_pos < 0.75).
+     * Letting conviction also drive safety unblocks Full realisation for
+     * confident policy actions at the smoke horizon, instead of requiring
+     * health graduation which 20-epoch smokes structurally can't reach. */
     if (!is_hold_action) {
         float h = fminf(1.0f, fmaxf(0.0f, health_for_kelly));
-        float safety = 0.5f + 0.5f * h;
+        float health_safety = 0.5f + 0.5f * h;
+        float conviction_clamped = fminf(1.0f, fmaxf(0.0f, conviction));
+        float safety = fmaxf(health_safety, conviction_clamped);
         target_position = apply_kelly_cap(
             target_position,
             *win_count, *loss_count,