diff --git a/ml/eval/eval_cl.py b/ml/eval/eval_cl.py
index 69133fc..695b363 100644
--- a/ml/eval/eval_cl.py
+++ b/ml/eval/eval_cl.py
@@ -54,9 +54,17 @@ def run(model, kin, tasks, device, bs=256, **kw):
         out.update({'reverse_p50': median(rr), 'reverse_mean': rr.mean().item(),
                     'reverse_moving_p50': median(flat['reverse_rate_moving']),   # 只在动了的步上算，不被刹停稀释
                     'moving_share_p50': median(flat['moving_share'])})          # 活动窗内动了的步占比（=1−站住比例）
-    if 'brake_rate' in flat:   # §9.2 判据是"刹停 <2% 推理周期"，所以要按规划次数归一，不能只报绝对次数
-        out.update({'brake_rate_mean': flat['brake_rate'].mean().item(), 'brake_rate_p50': median(flat['brake_rate']),
+    if 'brake_rate' in flat:   # §9.2 要"刹停 <2% 推理周期"。均值会被长尾拖着（实测 p50=0.5% 而 mean=10%），
+        rr = flat['brake_rate']  # 所以必须分层报：p90 与"卡住题占比"才是能区分杠杆的读数。
+        out.update({'brake_rate_mean': rr.mean().item(), 'brake_rate_p50': median(rr), 'brake_rate_p90': rr.quantile(0.9).item(),
+                    'stuck_share': (rr > 0.2).float().mean().item(),
                     'n_plans_mean': flat['n_plans'].mean().item()})
+    cats = [t.get('cat') for t in tasks]           # 分层读数：D 类按定义不可达，混在分母里会让"可达任务到达率"失真
+    reach = torch.tensor([c != 'D' for c in cats])
+    if reach.any():
+        out['success_reachable'] = (ok[reach].float().mean().item())
+    out['by_cat'] = {c: [int(sum(1 for x in cats if x == c)), int(sum(1 for i, x in enumerate(cats) if x == c and ok[i]))]
+                     for c in sorted({c for c in cats if c})}
     if 'backoff' in flat and flat['backoff'].any():   # backoff=缩步顶上的次数；brakes 现在只记连最小档都不安全的
         out['backoff_mean'] = flat['backoff'].mean().item()      # ⇒ 原来的"全拒"事件数 = 两者之和
         out['reject_all_mean'] = (flat['backoff'] + flat['brakes']).mean().item()