diff --git a/crates/ml/src/cuda_pipeline/gpu_dqn_trainer.rs b/crates/ml/src/cuda_pipeline/gpu_dqn_trainer.rs
index 3d4850d0b..6951daecb 100644
--- a/crates/ml/src/cuda_pipeline/gpu_dqn_trainer.rs
+++ b/crates/ml/src/cuda_pipeline/gpu_dqn_trainer.rs
@@ -2148,10 +2148,26 @@ impl GpuDqnTrainer {
         // Floor = max(10 * q_gap, 3 * q_std) — enough atoms to resolve actions.
         // With 51 atoms, a width of 10*q_gap gives q_gap/delta_z ≈ 5 atoms of
         // resolution between best and worst action.
+        //
+        // Clamp the adaptive [v_min, v_max] to the theoretical bounds supplied
+        // at construction (config.v_{min,max}, derived from reward_scale/gamma).
+        // Without this clamp the window can drift arbitrarily far from 0: once
+        // the policy overestimates, eval_q_mean_ema tracks it up → v_max follows
+        // → C51 atoms span a wider range → even larger TD targets → loss
+        // explodes geometrically (Fold 1 smoke: final_loss=1.6e16, grad=62B).
+        //
+        // Two clamps are needed — one on half-width so the band can't exceed
+        // the full theoretical range, and one on the centre so the band always
+        // fits inside [config.v_min, config.v_max]:
         let gap_width = (10.0 * q_gap).max(3.0 * self.eval_q_std_ema).max(0.1);
-        let half = gap_width;
-        let v_min = self.eval_q_mean_ema - half;
-        let v_max = self.eval_q_mean_ema + half;
+        let abs_half  = (self.config.v_max - self.config.v_min) * 0.5;
+        let half      = gap_width.min(abs_half);
+        let center    = self.eval_q_mean_ema.clamp(
+            self.config.v_min + half,
+            self.config.v_max - half,
+        );
+        let v_min = center - half;
+        let v_max = center + half;
         // Pinned device-mapped: CPU writes directly, GPU reads via dev_ptr. No HtoD copy.
         unsafe {
             *self.eval_v_range_pinned = v_min;