diff --git a/crates/ml-alpha/examples/phase1d_long_horizon.rs b/crates/ml-alpha/examples/phase1d_long_horizon.rs
index 32c40c1cc..0252e3955 100644
--- a/crates/ml-alpha/examples/phase1d_long_horizon.rs
+++ b/crates/ml-alpha/examples/phase1d_long_horizon.rs
@@ -36,7 +36,7 @@ use ml_alpha::fxcache_reader::{FxCacheReader, COL_RAW_CLOSE, FEAT_DIM};
 use ml_alpha::mamba2_block::{
     Mamba2AdamW, Mamba2AdamWConfig, Mamba2Block, Mamba2BlockConfig,
 };
-use ml_alpha::metrics_detail::{brier_score, log_loss};
+use ml_alpha::metrics_detail::{brier_score, log_loss, stratified_accuracy};
 use ml_alpha::multi_horizon_labels::generate_labels;
 
 use ml_core::cuda_autograd::gpu_tensor::GpuTensor;
@@ -305,6 +305,50 @@ fn main() -> Result<()> {
                  iso_acc, iso_auc, iso_brier, iso_logl);
     }
 
+    // ── Stratified accuracy on Block-S features (regime diagnostic) ──
+    //
+    // Block-S sits at cols 75..81 of the 81-dim snapshot row. For each
+    // val sequence, gather the END-BAR's feature value, then stratify
+    // val accuracy across quintiles of that feature. This tells us
+    // whether Mamba's K=6000 alpha is regime-conditional (concentrated
+    // in spread-Q4 / micro-drift-Q4 like the Phase 1c stateless MLP) or
+    // uniform — which decides whether we need an explicit regime head
+    // before the backtest.
+    const BLOCK_S_OFF: usize = 75;
+    const BLOCK_S_NAMES: [&str; 6] = [
+        "time_since_trade_s",
+        "time_since_snap_s",
+        "book_event_rate_per_s",
+        "spread_bps",
+        "L1_imbalance",
+        "micro_mid_drift",
+    ];
+    let labels_u8_full: Vec<u8> =
+        val_ys.iter().map(|&y| if y > 0.5 { 1 } else { 0 }).collect();
+    let val_labels_f32: Vec<f32> = labels_u8_full.iter().map(|&y| y as f32).collect();
+    println!();
+    println!("--- Stratified val accuracy across Block-S features (5 quintiles) ---");
+    for (col_off, name) in BLOCK_S_NAMES.iter().enumerate() {
+        let global_col = BLOCK_S_OFF + col_off;
+        let feat: Vec<f32> = val_pos.iter()
+            .map(|&p| {
+                let end_bar = labels.valid_indices[p];
+                feature_matrix[end_bar * alpha_dim + global_col]
+            })
+            .collect();
+        let strats = stratified_accuracy(&val_logits, &val_labels_f32, &feat, 5);
+        println!();
+        println!(" ▸ feature: {}", name);
+        println!("   {:>4} {:>14} {:>14} {:>10} {:>10} {:>12}",
+                 "k", "feat_lo", "feat_hi", "n", "accuracy", "observed");
+        for s in &strats {
+            println!(
+                "   {:>4} {:>14.4} {:>14.4} {:>10} {:>10.4} {:>12.4}",
+                s.stratum, s.feature_lo, s.feature_hi, s.n, s.accuracy, s.observed_pos_rate
+            );
+        }
+    }
+
     println!();
     if auc > 0.55 {
         println!("GATE PASS: AUC > 0.55 at K={}; multi-minute alpha confirmed.", cli.horizon);