diff --git a/crates/ml/src/cuda_pipeline/experience_kernels.cu b/crates/ml/src/cuda_pipeline/experience_kernels.cu
index 5bc144bf6..069c10c47 100644
--- a/crates/ml/src/cuda_pipeline/experience_kernels.cu
+++ b/crates/ml/src/cuda_pipeline/experience_kernels.cu
@@ -718,7 +718,8 @@ extern "C" __global__ void experience_action_select(
     float eps_ord_mult,    /* per-branch epsilon multiplier: order type */
     float eps_urg_mult,    /* per-branch epsilon multiplier: urgency */
     int timestep,          /* current timestep for stateless RNG */
-    const float* __restrict__ per_sample_epsilon  /* [N] IQL expectile gap epsilon, NULL=use cosine schedule */
+    const float* __restrict__ per_sample_epsilon,  /* [N] IQL expectile gap epsilon, NULL=use cosine schedule */
+    const float* __restrict__ isv_signals_ptr      /* [8] pinned ISV signals for adaptive hold. NULL = static. */
 ) {
     int i = blockIdx.x * blockDim.x + threadIdx.x;
     if (i >= N) return;
@@ -773,7 +774,16 @@ extern "C" __global__ void experience_action_select(
         int ps_base = i * 23;  /* PORTFOLIO_STRIDE = 23 */
         float hold_time_val = portfolio_states[ps_base + 10];
         cur_position = portfolio_states[ps_base + 0];
-        in_hold = (hold_time_val > 0.0f && hold_time_val < (float)min_hold_bars);
+        /* Adaptive hold: base + ISV-driven extension (same formula as Layer 2) */
+        int adaptive_hold = min_hold_bars;
+        if (isv_signals_ptr != NULL) {
+            float stability = 1.0f / (1.0f + isv_signals_ptr[1]);
+            float confidence = 1.0f - fminf(isv_signals_ptr[3], 1.0f);
+            int extension = (int)(stability * confidence * 8.0f);
+            if (isv_signals_ptr[5] < -0.001f) extension = 0;
+            adaptive_hold = min_hold_bars + extension;
+        }
+        in_hold = (hold_time_val > 0.0f && hold_time_val < (float)adaptive_hold);
     }
 
     if (in_hold) {
diff --git a/crates/ml/src/cuda_pipeline/gpu_experience_collector.rs b/crates/ml/src/cuda_pipeline/gpu_experience_collector.rs
index da04bd3e0..88bec88d5 100644
--- a/crates/ml/src/cuda_pipeline/gpu_experience_collector.rs
+++ b/crates/ml/src/cuda_pipeline/gpu_experience_collector.rs
@@ -2217,6 +2217,7 @@ impl GpuExperienceCollector {
                     .arg(&config.eps_urg_mult)          // per-branch epsilon multiplier: urgency
                     .arg(&(t as i32))                   // timestep for stateless Philox RNG
                     .arg(&self.per_sample_epsilon_ptr)   // IQL expectile gap epsilon (0=cosine schedule)
+                    .arg(&self.isv_signals_dev_ptr)      // ISV signals for adaptive hold (0=NULL=static)
                     .launch(launch_cfg)
                     .map_err(|e| MLError::ModelError(format!(
                         "experience_action_select t={t}: {e}"