diff --git a/docs/experiment_round38_backoff1.json b/docs/experiment_round38_backoff1.json
new file mode 100644
index 0000000..bf69d62
--- /dev/null
+++ b/docs/experiment_round38_backoff1.json
@@ -0,0 +1,84 @@
+{
+  "config": {
+    "ckpt": "ml/runs/base_dag2_dag1/best.pt",
+    "samples": "ml/data/samples",
+    "tasks": null,
+    "split": "val",
+    "n": 256,
+    "K": 8,
+    "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": null,
+    "out": "docs/experiment_round38_backoff1.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": 1,
+    "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": 0.52734375,
+    "collision_free_success": 0.52734375,
+    "time_ok_1p3": 0.26953125,
+    "collision_rate": 0.0,
+    "collision_steps_mean": 0.0,
+    "travel_ratio_median": 2.615795850753784,
+    "time_mean_s": 2.479703664779663,
+    "brakes_mean": 18.77734375,
+    "start_fail": 0.0,
+    "goal_ratio_2s": 0.27920001745224,
+    "err_fail_med": 212.4702606201172,
+    "err_success_med": 0.20242473483085632,
+    "dls_share": 0.6171875,
+    "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": 0.2482638955116272,
+    "reverse_mean": 0.2603996992111206,
+    "reverse_moving_p50": 0.3252427279949188,
+    "moving_share_p50": 0.9640287756919861,
+    "brake_rate_mean": 0.09629407525062561,
+    "brake_rate_p50": 0.0051282052882015705,
+    "n_plans_mean": 195.0,
+    "backoff_mean": 1.57421875,
+    "reject_all_mean": 20.3515625
+  },
+  "simulation_wall_s": 334.8870786000043
+}
diff --git a/docs/experiment_round38_backoff2.json b/docs/experiment_round38_backoff2.json
new file mode 100644
index 0000000..9f44646
--- /dev/null
+++ b/docs/experiment_round38_backoff2.json
@@ -0,0 +1,84 @@
+{
+  "config": {
+    "ckpt": "ml/runs/base_dag2_dag1/best.pt",
+    "samples": "ml/data/samples",
+    "tasks": null,
+    "split": "val",
+    "n": 256,
+    "K": 8,
+    "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": null,
+    "out": "docs/experiment_round38_backoff2.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": 2,
+    "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": 0.5234375,
+    "collision_free_success": 0.5234375,
+    "time_ok_1p3": 0.26171875,
+    "collision_rate": 0.0,
+    "collision_steps_mean": 0.0,
+    "travel_ratio_median": 2.5341429710388184,
+    "time_mean_s": 2.4558210372924805,
+    "brakes_mean": 18.1171875,
+    "start_fail": 0.0,
+    "goal_ratio_2s": 0.274954229593277,
+    "err_fail_med": 207.97683715820312,
+    "err_success_med": 0.19538111984729767,
+    "dls_share": 0.60546875,
+    "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": 0.24827586114406586,
+    "reverse_mean": 0.2605503499507904,
+    "reverse_moving_p50": 0.32248520851135254,
+    "moving_share_p50": 0.9672130942344666,
+    "brake_rate_mean": 0.09290865808725357,
+    "brake_rate_p50": 0.0051282052882015705,
+    "n_plans_mean": 195.0,
+    "backoff_mean": 4.99609375,
+    "reject_all_mean": 23.11328125
+  },
+  "simulation_wall_s": 369.2254191000102
+}
diff --git a/docs/experiment_round38_backoff3.json b/docs/experiment_round38_backoff3.json
new file mode 100644
index 0000000..d0168cc
--- /dev/null
+++ b/docs/experiment_round38_backoff3.json
@@ -0,0 +1,84 @@
+{
+  "config": {
+    "ckpt": "ml/runs/base_dag2_dag1/best.pt",
+    "samples": "ml/data/samples",
+    "tasks": null,
+    "split": "val",
+    "n": 256,
+    "K": 8,
+    "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": null,
+    "out": "docs/experiment_round38_backoff3.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": 3,
+    "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": 0.515625,
+    "collision_free_success": 0.515625,
+    "time_ok_1p3": 0.2578125,
+    "collision_rate": 0.0,
+    "collision_steps_mean": 0.0,
+    "travel_ratio_median": 2.529658794403076,
+    "time_mean_s": 2.4427273273468018,
+    "brakes_mean": 16.48046875,
+    "start_fail": 0.0,
+    "goal_ratio_2s": 0.2868920564651489,
+    "err_fail_med": 208.72935485839844,
+    "err_success_med": 0.20242473483085632,
+    "dls_share": 0.60546875,
+    "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": 0.25,
+    "reverse_mean": 0.26120537519454956,
+    "reverse_moving_p50": 0.3236245810985565,
+    "moving_share_p50": 0.9672130942344666,
+    "brake_rate_mean": 0.08451522886753082,
+    "brake_rate_p50": 0.0051282052882015705,
+    "n_plans_mean": 195.0,
+    "backoff_mean": 9.703125,
+    "reject_all_mean": 26.18359375
+  },
+  "simulation_wall_s": 416.57447530000354
+}
diff --git a/docs/探索日志.md b/docs/探索日志.md
index f025c90..1056045 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。
 
+## `backoff` 四臂判掉：转化率能买到刹停率，买不到到达——而且它能**绕过 §9.2 的判据**（轮次 43）
+
+- 同一 256 题同 seed、`--inc_cap 1 --cert_batch 1`，只改缩步档数（`docs/experiment_round38_backoff{0,1,2,3}.json`）：
+
+  | 档数 | 到达 | 时限内 | 刹停/题 | 缩步顶上/题 | 全拒事件合计 | 转化率 | `brake_rate` mean |
+  |---|---|---|---|---|---|---|---|
+  | 0（对照） | 0.5234 | 0.2656 | 19.55 | 0 | — | — | **10.03%** |
+  | 1（1/2） | 0.5273 | 0.2695 | 18.78 | 1.57 | 20.35 | **7.7%** | 9.63% |
+  | 2（+1/4） | 0.5234 | 0.2617 | 18.12 | 5.00 | 23.11 | 21.6% | 9.29% |
+  | 3（+1/8） | **0.5156** | **0.2578** | 16.48 | 9.70 | **26.18** | 37.1% | **8.45%** |
+
+- **预注册判读命中低那一支**（转化率 <60%）：把全拒变成"慢一拍的动作"确实随档数上升（7.7%→37.1%），但**到达和时限一点没赚**（0.5234/0.2656 → 0.5156/0.2578，方向还是负的）。⇒ 被全拒的那些状态不是"差一点就能过"，而是**这个方向本来就不通**：缩步只是慢一点去撞同一堵墙。
+- 更值得警惕的是第二列和最后一列的**背离**：`brake_rate`（§9.2 要 <2% 的那个判据）从 10.03% 一路降到 8.45%，`brakes_mean` 降 16%，**但真实结果没变好**。也就是说**缩步档可以只改善刹停率指标而不改善行为**——判据本身能被绕过。而且"全拒事件合计"反而从 20.35 涨到 26.18：缩步把机械臂带进了更容易被拒的几何里。
+  ⇒ 规则：**任何降低刹停率的改动，必须同时报到达与时限**，只看 `brake_rate` 会奖励错误行为。这条已加进 #25 的分层读数要求。
+- 处置：`backoff` 保持默认关、**作为到达率杠杆判掉**；它在"减少原地站住"这件事上确实有效（−16% 刹停），所以留作实机上"不许完全停住"这类需求的安全侧选项，不再投入档位扫描。
+- 下一步（链已结束，可以改库）：#25 分层指标（可达分母 / 按类别到达率 / `brake_rate_p90` / `stuck_share`）→ 用它重读 K=16（#27：+3.9 pp 到底买在哪一类）→ 才轮到 #24（几何信息进打分，直接打 B/C 类"方向不通"）。
+
 ## 控制臂对表：22/22 键逐位复现 + 仪表预测精确命中（轮次 42）
 
 - `--backoff 0 --cert_batch 1 --inc_cap 1` 重跑轮次 33 cap 配方：**两文件共有的 22 个指标键全部相等，差异集合为空**（第三次确认批量认证分支重写没有改变行为）。