From f42d3d756f6517033f86da93710a1b89e8483eec Mon Sep 17 00:00:00 2001
From: cbb <2625274797@qq.com>
Date: Tue, 20 Oct 2026 21:57:02 +0800
Subject: [PATCH] =?UTF-8?q?exp:=20=E8=BD=AE=E6=AC=A138(=E8=A1=A5)=E2=80=94?=
 =?UTF-8?q?=E2=80=94steps=3D4=20=E6=9C=AB=E6=80=81=E8=AF=AF=E5=B7=AE0.135m?=
 =?UTF-8?q?m=E6=9C=80=E4=BC=98=E4=BD=86=E4=BB=8D=E8=A2=AB=E6=8B=9280?=
 =?UTF-8?q?=E6=AC=A1/=E9=A2=98=EF=BC=9B=E5=A4=B1=E8=B4=A5=E5=88=86?=
 =?UTF-8?q?=E7=B1=BB=E8=AF=81=E6=98=8E=E7=BC=BA=E5=8F=A3=E6=98=AF=E8=B5=B0?=
 =?UTF-8?q?=E4=B8=80=E5=8D=8A=E8=A2=AB=E5=86=BB=E4=BD=8F(63/109)?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 docs/experiment_round36_steps4_cap1.json   | 73 ++++++++++++++++++++++
 docs/experiment_round38_fail_taxonomy.json | 19 ++++++
 docs/探索日志.md                           | 17 ++++-
 3 files changed, 108 insertions(+), 1 deletion(-)
 create mode 100644 docs/experiment_round36_steps4_cap1.json
 create mode 100644 docs/experiment_round38_fail_taxonomy.json

diff --git a/docs/experiment_round36_steps4_cap1.json b/docs/experiment_round36_steps4_cap1.json
new file mode 100644
index 0000000..7f82067
--- /dev/null
+++ b/docs/experiment_round36_steps4_cap1.json
@@ -0,0 +1,73 @@
+{
+  "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": "ml/runs/round36_steps4_cap1_rows.json",
+    "out": "docs/experiment_round36_steps4_cap1.json",
+    "trace": null,
+    "select": "endpoint",
+    "replan": "empty",
+    "safety": 1,
+    "partial": 1,
+    "inc_cap": 1,
+    "gap_margin": 0.0,
+    "cert_batch": 0,
+    "steps": 4,
+    "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": 0.40625,
+    "collision_free_success": 0.40625,
+    "time_ok_1p3": 0.3046875,
+    "collision_rate": 0.0,
+    "collision_steps_mean": 0.0,
+    "travel_ratio_median": 1.8817013502120972,
+    "time_mean_s": 1.9080770015716553,
+    "brakes_mean": 79.65234375,
+    "start_fail": 0.0,
+    "goal_ratio_2s": 0.3844277262687683,
+    "err_fail_med": 342.1692810058594,
+    "err_success_med": 0.1346309632062912,
+    "dls_share": 0.4609375,
+    "time_budget": "teacher×2(≥3 s)",
+    "speed_ratio_p50": 1.0000016689300537,
+    "speed_ratio_p95": 1.0000053644180298,
+    "speed_ratio_max": 1.0000101327896118,
+    "over_vmax_rate": 0.0,
+    "accel_ratio_p95": 0.914295494556427,
+    "accel_ratio_max": 1.0000005960464478
+  },
+  "simulation_wall_s": 347.79689209999924
+}
diff --git a/docs/experiment_round38_fail_taxonomy.json b/docs/experiment_round38_fail_taxonomy.json
new file mode 100644
index 0000000..d158e93
--- /dev/null
+++ b/docs/experiment_round38_fail_taxonomy.json
@@ -0,0 +1,19 @@
+[
+ {
+  "label": "steps=1 cap=0",
+  "n": 256,
+  "arrived": 147,
+  "failed": 109,
+  "buckets": {
+   "姿态没到位(位置<5mm)": 3,
+   "位置 5~20mm": 5,
+   "位置 20~100mm": 15,
+   "位置 >100mm 且刹车>30(被挡)": 63,
+   "位置 >100mm 且刹车<=30(方向/绕圈)": 23,
+   "起步失败": 0
+  },
+  "failed_e_p50": 255.333598802122,
+  "failed_oe_p50": 11.016554443385107,
+  "failed_still_share_p50": 0.3722222149372101
+ }
+]
\ No newline at end of file
diff --git a/docs/探索日志.md b/docs/探索日志.md
index c3337e6..7df016b 100644
--- a/docs/探索日志.md
+++ b/docs/探索日志.md
@@ -34,8 +34,23 @@
   | steps=2 cap=0 | 0.4258 | 0.3555 | **86.31** | 1.67 s | **1.571** | 0.243 mm | 0.453 |
   | steps=1 cap=1（=轮次33 cap） | 0.5234 | 0.2656 | 19.55 | 2.49 s | 2.616 | 0.218 mm | 0.613 |
   | steps=2 cap=1 | 0.3945 | 0.2969 | 86.16 | 1.93 s | 1.844 | 0.293 mm | 0.445 |
+  | steps=4 cap=1 | 0.4062 | 0.3047 | 79.65 | 1.91 s | 1.882 | **0.135 mm** | 0.461 |
 
-  （`docs/experiment_round36_steps2_cap{0,1}.json`；steps=4/8 同批续跑）
+  （`docs/experiment_round36_steps2_cap{0,1}.json`、`docs/experiment_round36_steps4_cap1.json`；steps=8 同批续跑）
+  段数再多并不是越差：steps=4 比 steps=2 到达与时限都略升，成功题末态误差 0.293 → **0.135 mm**（比一步口径的 0.218 还好）——但刹车仍近 80/题，到达率被同一道门卡住。
+- **失败分类**（不占卡：读轮次 33 base 的 256 题轨迹转储，CPU 重算末态位姿误差，`docs/experiment_round38_fail_taxonomy.json`）。147 到达 / 109 未到，未到题按"末态离目标多远 + 被拒多少次"分桶：
+
+  | 桶 | 题数 |
+  |---|---|
+  | 位置 >100 mm 且刹车 >30 次（走一半被冻住） | **63** |
+  | 位置 >100 mm 但刹车 ≤30 次（方向错 / 绕圈） | 23 |
+  | 位置 20~100 mm | 15 |
+  | 位置 5~20 mm | 5 |
+  | 位置 <5 mm 但姿态没到位 | 3 |
+  | 起步失败 | 0 |
+
+  未到组中位数：位置误差 255 mm、姿态误差 11.0°、**活动窗内 37% 的步是零运动**。
+  ⇒ 到达率的缺口不在精度或姿态判据（这两类合计只有 8 题），而在**走一半被冻住**：76% 的失败离目标还 >100 mm，其中多数是安全层反复全拒。这正是 `backoff`（全拒时缩步顶上）对着的那块；也说明 #15（候选数与安全档位）比"往打分里加剩余时限"更值钱。
 - **反直觉但自洽的一组数**：两段推理**开环更贴老师**（E1-b：0.43 vs 4.33）、**行程比更接近 1**（1.571 vs 2.445）、**到了的题还更快**（1.67 vs 2.31 s），但闭环到达掉 14.8 个百分点。逐题对照：108 题两者都到、**39 题只有 1 步能到、只有 1 题是 2 步独有** ⇒ 不是"换了一批题"，是同一批题里 1 步原本能做对的那部分被卡住了。
 - **卡点在安全层不在策略**：刹车从 19.25 → 86.3/题（4.5 倍），未到达组的刹车中位数从 45 → 176（几乎每步都被拒 ⇒ 原地冻结，末态误差 p50 也从 230 → 406 mm）。也就是说**发布的那 57.4% 是与"一步那套偏置且被削弱过的动作"共同标定出来的**：策略、反馈增益、认证口径、DAgger 采集分布四者都围着它。
 - **两种解释必须分开，否则下一步会走错**：