diff --git a/crates/ml/src/cuda_pipeline/gpu_dqn_trainer.rs b/crates/ml/src/cuda_pipeline/gpu_dqn_trainer.rs
index 7abb2bc55..cdc9f0ca3 100644
--- a/crates/ml/src/cuda_pipeline/gpu_dqn_trainer.rs
+++ b/crates/ml/src/cuda_pipeline/gpu_dqn_trainer.rs
@@ -2153,6 +2153,24 @@ impl GpuDqnTrainer {
             let out_ptr = atom_positions_buf_ptr + out_offset;
             let branch_i32 = branch as i32;
 
+            // ISV v-range diag: log host-side ISV read used by the kernel.
+            // The kernel reads ISV via isv_dev_ptr (device mirror); the
+            // pinned mirror tracks the same writes from update_eval_v_range.
+            let (diag_center, diag_half) = if !self.isv_signals_pinned.is_null() {
+                unsafe {
+                    let c = *self.isv_signals_pinned.add(V_CENTER_DIR_INDEX + 2 * branch);
+                    let h = *self.isv_signals_pinned.add(V_CENTER_DIR_INDEX + 2 * branch + 1);
+                    (c, h)
+                }
+            } else {
+                (f32::NAN, f32::NAN)
+            };
+            tracing::info!(
+                target: "isv_vrange_diag",
+                branch, diag_center, diag_half,
+                "recompute_atom_positions ISV read"
+            );
+
             unsafe {
                 self.stream.launch_builder(&self.adaptive_atom_kernel)
                     .arg(&spacing_ptr)
@@ -2211,6 +2229,12 @@ impl GpuDqnTrainer {
                 if (b_max - b_min).abs() < 1e-6 { (v_min_f, v_max_f) }
                 else { (b_min, b_max) }
             };
+            tracing::info!(
+                target: "isv_vrange_diag",
+                branch, b_min, b_max,
+                isv_ptr_null = isv_ptr.is_null(),
+                "warm_start clamp bounds"
+            );
             let slot = &mut host_data[branch * na..(branch + 1) * na];
             for (atom_idx, q) in quantiles.iter().enumerate() {
                 slot[atom_idx] = q.clamp(b_min, b_max);
@@ -2220,6 +2244,21 @@ impl GpuDqnTrainer {
         self.stream.memcpy_htod(&host_data, &mut self.atom_positions_buf)
             .map_err(|e| MLError::ModelError(format!("warm_start_atom_positions HtoD: {e}")))?;
 
+        // ISV v-range diag: log what we just wrote per branch (min/max of
+        // atom_positions) — host_data is the source of truth (no need for a
+        // DtoH roundtrip since memcpy_htod doesn't mutate source on failure
+        // and succeeded above).
+        for branch in 0..4_usize {
+            let slot = &host_data[branch * na..(branch + 1) * na];
+            let amin = slot.iter().cloned().fold(f32::INFINITY, f32::min);
+            let amax = slot.iter().cloned().fold(f32::NEG_INFINITY, f32::max);
+            tracing::info!(
+                target: "isv_vrange_diag",
+                branch, amin, amax,
+                "warm_start atom_positions written"
+            );
+        }
+
         Ok(())
     }
 
@@ -3058,6 +3097,10 @@ impl GpuDqnTrainer {
             let q_mean = stats.q_mean;
             let q_std = stats.q_variance.max(0.0).sqrt();
             let q_gap = per_branch_q_gaps[branch_idx].max(0.0);
+            // ISV v-range diag: capture the branch's initialization state
+            // BEFORE the EMA update so the log reflects which cold/warm path
+            // produced the (center, half) being written.
+            let initialized_before = self.eval_ema_initialized[branch_idx];
 
             // Adaptive-rate EMA with baseline proportional to Q-std.
             if !self.eval_ema_initialized[branch_idx] {
@@ -3090,6 +3133,13 @@ impl GpuDqnTrainer {
                 v_max_f - half,
             );
 
+            tracing::info!(
+                target: "isv_vrange_diag",
+                branch_idx, q_mean, q_std, q_gap, center, half,
+                initialized = initialized_before,
+                "update_eval_v_range branch"
+            );
+
             unsafe {
                 *self.isv_signals_pinned.add(V_CENTER_DIR_INDEX + 2 * branch_idx) = center;
                 *self.isv_signals_pinned.add(V_CENTER_DIR_INDEX + 2 * branch_idx + 1) = half;