diff --git a/crates/ml/src/cuda_pipeline/gpu_dqn_trainer.rs b/crates/ml/src/cuda_pipeline/gpu_dqn_trainer.rs
index 71e43d12f..d082d4494 100644
--- a/crates/ml/src/cuda_pipeline/gpu_dqn_trainer.rs
+++ b/crates/ml/src/cuda_pipeline/gpu_dqn_trainer.rs
@@ -542,6 +542,8 @@ pub struct GpuDqnTrainer {
     iqn_trunk_adam_step: i32,
     /// IQN trunk grad norm scratch [1] (f32 accumulator).
     iqn_trunk_grad_norm: CudaSlice<f32>,
+    /// Scratch bf16 [1] for IQN/ensemble finalize — prevents overwriting main grad_norm_buf.
+    aux_norm_bf16_scratch: CudaSlice<half::bf16>,
     /// IQN trunk Adam step counter on device [1].
     iqn_trunk_t_buf: CudaSlice<i32>,
     /// Number of trunk parameters (w_s1 + b_s1 + w_s2 + b_s2).
@@ -1217,7 +1219,8 @@ impl GpuDqnTrainer {
             let scratch_ptr = self.ptrs.iqn_trunk_m;
             let partials_ptr = self.grad_norm_partials.raw_ptr();
             let norm_ptr = self.ptrs.iqn_trunk_grad_norm;
-            let bf16_ptr = self.grad_norm_buf.raw_ptr();
+            // Use dedicated scratch for bf16 output — NOT main grad_norm_buf
+            let bf16_ptr = self.aux_norm_bf16_scratch.raw_ptr();
             let n_i32 = trunk_grad_total as i32;
             let blocks = ((trunk_grad_total + 255) / 256) as u32;
             let nb = blocks as i32;
@@ -1235,7 +1238,7 @@ impl GpuDqnTrainer {
                     })
                     .map_err(|e| MLError::ModelError(format!("IQN trunk grad_norm phase1: {e}")))?;
             }
-            // Phase 2: reduce → iqn_trunk_grad_norm
+            // Phase 2: reduce → iqn_trunk_grad_norm (f32) + scratch (bf16)
             unsafe {
                 self.stream
                     .launch_builder(&self.grad_norm_finalize_kernel)
@@ -1477,7 +1480,7 @@ impl GpuDqnTrainer {
             let scratch_ptr = self.ptrs.iqn_trunk_m;
             let partials_ptr = self.grad_norm_partials.raw_ptr();
             let norm_ptr = self.ptrs.iqn_trunk_grad_norm;
-            let bf16_ptr = self.grad_norm_buf.raw_ptr();
+            let bf16_ptr = self.aux_norm_bf16_scratch.raw_ptr();
             let n_i32 = trunk_grad_total as i32;
             let blocks = ((trunk_grad_total + 255) / 256) as u32;
             let nb = blocks as i32;
@@ -2507,6 +2510,8 @@ impl GpuDqnTrainer {
         let iqn_trunk_v = alloc_bf16(&stream, trunk_params, "iqn_trunk_v")?;
         let iqn_trunk_grad_norm = stream.alloc_zeros::<f32>(1)
             .map_err(|e| MLError::ModelError(format!("alloc iqn_trunk_grad_norm f32: {e}")))?;
+        // Scratch bf16 for IQN/ensemble trunk finalize — prevents overwriting main grad_norm_buf.
+        let aux_norm_bf16_scratch = alloc_bf16(&stream, 1, "aux_norm_bf16_scratch")?;
         let iqn_trunk_t_buf = alloc_i32(&stream, 1, "iqn_trunk_t_buf")?;
         let mut spec_u_s1 = alloc_bf16(&stream, config.shared_h1, "spec_u_s1")?;
         let mut spec_v_s1 = alloc_bf16(&stream, config.state_dim, "spec_v_s1")?;
@@ -2915,6 +2920,7 @@ impl GpuDqnTrainer {
             iqn_trunk_v,
             iqn_trunk_adam_step: 0,
             iqn_trunk_grad_norm,
+            aux_norm_bf16_scratch,
             iqn_trunk_t_buf,
             trunk_param_count: trunk_params,
             states_buf,