diff --git a/crates/ml-dqn/src/gpu_replay_buffer.rs b/crates/ml-dqn/src/gpu_replay_buffer.rs
index b3dc766ac..010d4e8b1 100644
--- a/crates/ml-dqn/src/gpu_replay_buffer.rs
+++ b/crates/ml-dqn/src/gpu_replay_buffer.rs
@@ -595,13 +595,15 @@ impl GpuReplayBuffer {
         // then index_add the deltas back. Single batched GPU kernel, no CPU loop.
         //
         // For rare duplicate indices (batch_size << capacity), the deltas may
-        // stack incorrectly, but this is benign: priorities will be slightly off
-        // and self-correct on the next update. This matches the old CPU loop's
-        // "last-write-wins" semantics in expectation.
+        // stack — index_add adds ALL deltas, so duplicates double-apply.
+        // Post-clamp prevents priorities from going to zero/negative (→ NaN weights).
         let indices_i64 = indices.to_dtype(DType::I64)?;
         let old_prios = self.priorities.index_select(&indices_i64, 0)?;
         let delta = (&clamped - &old_prios)?;
-        self.priorities = self.priorities.index_add(&indices_i64, &delta, 0)?;
+        self.priorities = self
+            .priorities
+            .index_add(&indices_i64, &delta, 0)?
+            .clamp(self.config.epsilon, 1e6)?;
 
         Ok(())
     }