diff --git a/crates/ml/src/cuda_pipeline/dqn_utility_kernels.cu b/crates/ml/src/cuda_pipeline/dqn_utility_kernels.cu
index b598e33ac..e4b25968a 100644
--- a/crates/ml/src/cuda_pipeline/dqn_utility_kernels.cu
+++ b/crates/ml/src/cuda_pipeline/dqn_utility_kernels.cu
@@ -1187,18 +1187,25 @@ extern "C" __global__ void bn_tanh_concat_f32_kernel(
  * ══════════════════════════════════════════════════════════════════════ */
 extern "C" __global__ void stochastic_depth_rng(
     float* __restrict__ scale_buf,         /* [3] output: per-layer scales */
-    unsigned int* __restrict__ rng_state,   /* [1] LCG state (read-write) */
+    unsigned int* __restrict__ rng_state,   /* [1] step counter (read-write) */
     float drop_prob                         /* probability of dropping each layer */
 ) {
-    unsigned int rng = rng_state[0];
+    /* Stateless Philox: deterministic per (step, layer). No accumulated LCG state. */
+    unsigned int step = rng_state[0];
     float keep_scale = 1.0f / (1.0f - drop_prob);
 
     for (int layer = 0; layer < 3; layer++) {
-        rng = rng * 1664525u + 1013904223u;
-        float u = (float)(rng & 0x00FFFFFFu) / 16777216.0f;
+        unsigned int key = step;
+        unsigned int ctr = (unsigned int)(layer * 37 + 7777);
+        ctr ^= key * 0x9E3779B9u;
+        ctr *= 0x85ebca6bu;
+        ctr ^= ctr >> 13;
+        ctr *= 0xc2b2ae35u;
+        ctr ^= ctr >> 16;
+        float u = (float)(ctr & 0x00FFFFFFu) / 16777216.0f;
         scale_buf[layer] = (u > drop_prob) ? keep_scale : 0.0f;
     }
-    rng_state[0] = rng;
+    rng_state[0] = step + 1; /* monotonic counter, not LCG */
 }
 
 /* ══════════════════════════════════════════════════════════════════════