diff --git a/crates/ml-alpha/cuda/rl_confidence_gate.cu b/crates/ml-alpha/cuda/rl_confidence_gate.cu
index bf40c024e..521c0c66d 100644
--- a/crates/ml-alpha/cuda/rl_confidence_gate.cu
+++ b/crates/ml-alpha/cuda/rl_confidence_gate.cu
@@ -89,9 +89,15 @@ extern "C" __global__ void rl_confidence_gate(
     }
     const float sigma = sqrtf(var + 1e-8f);
 
-    // Lower confidence bound.
-    const float lcb = mu - lambda * sigma;
-    const float conf = fmaxf(0.0f, fminf(lcb / (sigma_norm + 1e-8f), 1.0f));
+    // Lower confidence bound, shifted relative to atom span so conf=0
+    // means "worst possible" and conf=1 means "certain at V_MAX".
+    // Without the V_MIN shift, any distribution with σ > μ gives conf=0
+    // (permanent block for near-uniform Q at training start).
+    const float v_min = atom_supports[0];
+    const float v_max = atom_supports[Q_N_ATOMS - 1];
+    const float span = v_max - v_min + 1e-8f;
+    const float lcb = (mu - v_min) - lambda * sigma;
+    const float conf = fmaxf(0.0f, fminf(lcb / span, 1.0f));
 
     if (conf < threshold) {
         actions[b] = ACTION_HOLD;