diff --git a/docs/experiment_round42_K16cb1_dump.json b/docs/experiment_round42_K16cb1_dump.json
new file mode 100644
index 0000000..5fdd9e1
--- /dev/null
+++ b/docs/experiment_round42_K16cb1_dump.json
@@ -0,0 +1,107 @@
+{
+  "config": {
+    "ckpt": "ml/runs/base_dag2_dag1/best.pt",
+    "samples": "ml/data/samples",
+    "tasks": null,
+    "split": "val",
+    "n": 256,
+    "K": 16,
+    "commit": 2,
+    "fb": 2.2,
+    "ki": 0.0,
+    "fb_scale": 0.0,
+    "fb_near": 0.2,
+    "fb_fallback": 0.0,
+    "kdecay": 0.9,
+    "kimax": 0.5,
+    "fb_every": 1,
+    "safety_try": 0,
+    "sel_w": 0.1,
+    "dls_safety": 1,
+    "fb_clear": 0.0,
+    "fb_cmin": 0.3,
+    "sub_refresh": 0,
+    "fb_block_gain": 0.5,
+    "dump": "ml/runs/round42_K16cb1_rows.json",
+    "out": "docs/experiment_round42_K16cb1_dump.json",
+    "trace": null,
+    "select": "endpoint",
+    "replan": "empty",
+    "safety": 1,
+    "partial": 1,
+    "inc_cap": 1,
+    "gap_margin": 0.0,
+    "cert_batch": 1,
+    "steps": 1,
+    "z_scale": 1.0,
+    "len_w": 0.0,
+    "backoff": 0,
+    "max_sec": null,
+    "seed": 0,
+    "dls": "analytic",
+    "device": "cuda"
+  },
+  "batch_size": 256,
+  "batch_seed_rule": "same seed per batch",
+  "checkpoint_sha256": "f7f686fd8ee92fb6948fbcc4ee954636065ad87bb54f661c4f8f575bc16c2ce1",
+  "tasks_sha256": "249bd93457825cda98c59e64d9a096d3473c69aeb3553de1dea6d9e56be71772",
+  "kin_fp": "2ebcbf17",
+  "act_scale": 1.0,
+  "metrics": {
+    "n": 256,
+    "n_ref": 256,
+    "success": 0.5625,
+    "collision_free_success": 0.5625,
+    "time_ok_1p3": 0.296875,
+    "collision_rate": 0.0,
+    "collision_steps_mean": 0.0,
+    "travel_ratio_median": 2.6243183612823486,
+    "time_mean_s": 2.6024999618530273,
+    "brakes_mean": 13.7734375,
+    "start_fail": 0.0,
+    "goal_ratio_2s": 0.2799961566925049,
+    "err_fail_med": 195.448486328125,
+    "err_success_med": 0.15947841107845306,
+    "dls_share": 0.640625,
+    "time_budget": "teacher×2(≥3 s)",
+    "speed_ratio_p50": 1.0000016689300537,
+    "speed_ratio_p95": 1.0000101327896118,
+    "speed_ratio_max": 1.0000101327896118,
+    "over_vmax_rate": 0.0,
+    "accel_ratio_p95": 1.0000016689300537,
+    "accel_ratio_max": 1.0000101327896118,
+    "reverse_p50": 0.26249998807907104,
+    "reverse_mean": 0.2693643867969513,
+    "reverse_moving_p50": 0.3194444477558136,
+    "moving_share_p50": 0.9869565367698669,
+    "brake_rate_mean": 0.07063301652669907,
+    "brake_rate_p50": 0.0,
+    "brake_rate_p90": 0.2589743733406067,
+    "stuck_share": 0.15234375,
+    "n_plans_mean": 195.0,
+    "success_reachable": 0.5925925970077515,
+    "by_cat": {
+      "A": [
+        128,
+        100
+      ],
+      "B": [
+        74,
+        24
+      ],
+      "C": [
+        32,
+        12
+      ],
+      "D": [
+        13,
+        0
+      ],
+      "Dr": [
+        9,
+        8
+      ]
+    }
+  },
+  "simulation_wall_s": 613.4877918999991
+}
diff --git a/docs/探索日志.md b/docs/探索日志.md
index 70d0524..60b08c8 100644
--- a/docs/探索日志.md
+++ b/docs/探索日志.md
@@ -108,6 +108,23 @@
 - **对 `backoff` 的预期与判据**：全拒意味着"连第 1 步的端点或中点已经穿进障碍"，缩步能不能救取决于**穿得有多深**（穿 0.1 mm 救得回，穿 3 mm 救不回）。这个数转储里没记（只记被选中那条），所以直接由 `backoff_mean`/`reject_all_mean` 来量：
   `reject_all_mean ≈ 原全拒次数`（=backoff_mean+brakes_mean，与轮次 33 cap 的 19.55 可比），`backoff_mean/(backoff_mean+brakes_mean)` 就是"缩步能救回的比例"。这是补跑（#22）要读的第一个数。
 
+## #27：K=16 的增益**不集中在 B/C**——B 类缺口是策略能力，不是候选容量（轮次 44，逐题 dump）
+
+- K=8 与 K=16（都 `--cert_batch 1 --inc_cap 1`，同 256 题同 seed）逐题对齐后按类别：
+
+  | 类别 | 题数 | K8 到达 | K16 到达 | 差 | K8 刹停/题 | K16 刹停/题 |
+  |---|---|---|---|---|---|---|
+  | A 均匀 | 128 | 75% | 78% | +3.1 pp | 7.5 | 4.9 |
+  | B 直线被挡 | 74 | 27% | **32%** | +5.4 pp | 38.1 | **27.0** |
+  | C 贴立柱 | 32 | 34% | 38% | +3.1 pp | 33.3 | 24.5 |
+  | D 不可达 | 13 | 0% | 0% | 0 | 5.2 | 4.8 |
+  | Dr | 9 | 78% | 89% | +11.1 pp | 10.2 | 6.4 |
+
+  逐题翻转：**新救回 12 题（A 6、B 4、C 1、Dr 1），反而丢 2 题（都在 A）**，净 +10 题 = 总体 0.5234 → 0.5625。
+- **统计诚实**：分类切片里 1–4 题的变化在这个分母上不构成显著（B 类 +4 题 = +5.4 pp，单看会高估）；**稳的只有两条**——总体 +3.9 pp（净 10 题）与刹停整体 −29%（B 从 38.1 降到 27.0 次/题，是 74 题上的均值移动）。
+- **关键读法（决定预算投向）**：候选容量翻倍把 B 的**拒绝次数**砍掉近三成，但 B 的到达只从 27% 到 32%，离 A 类的 78% 还差 46 pp。⇒ **"少被拒"不等于"会绕"**：B/C 的缺口是策略在这些状态下**根本不产出绕行轨迹**，不是被闸门挡住的。继续在排序/认证口径上打磨救不了 B/C；要动的是数据侧（在受阻状态下采集教师轨迹，即 DAgger 第三轮，需要放行）或明确把 B/C 列为当前能力边界之外。
+- 由此更新候选清单：#24 修正版（`tie='gap'`，等长前缀时按沿路最小间隙定夺）仍然值得测——它影响的是"选了哪条活路"，会影响后续拍的存活率；但它**不能**被期待修好 B/C 的 46 pp，那属于训练侧。
+
 ## 分层指标上线并被三重交叉验证（轮次 43）
 
 - 对照配方（轮次 33 cap：1 步、`--inc_cap 1`）重跑一次带分层指标的评估：**原有 22 个指标逐位不变**，新增键全部产出（`docs/experiment_round43_control_strat.json`）：