From 1b301cc58eb69841c303bd8ca26a39c5eca9a688 Mon Sep 17 00:00:00 2001
From: cbb <2625274797@qq.com>
Date: Tue, 20 Oct 2026 23:51:51 +0800
Subject: [PATCH] =?UTF-8?q?exp:=20=E5=88=86=E5=B1=82=E6=8C=87=E6=A0=87?=
 =?UTF-8?q?=E4=B8=89=E9=87=8D=E4=BA=A4=E5=8F=89=E9=AA=8C=E8=AF=81=E9=80=9A?=
 =?UTF-8?q?=E8=BF=87=E2=80=94=E2=80=94success=5Freachable=200.5514/by=5Fca?=
 =?UTF-8?q?t(B27%,C34%)/brake=5Frate=20p50=200.5%=E4=BD=86p90=2031.5%?=
 =?UTF-8?q?=E3=80=81stuck=5Fshare=2023.4%=EF=BC=9B=E8=80=8122=E9=94=AE?=
 =?UTF-8?q?=E9=80=90=E4=BD=8D=E4=B8=8D=E5=8F=98?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 docs/experiment_round43_control_strat.json | 107 +++++++++++++++++++++
 docs/探索日志.md                           |  15 +++
 2 files changed, 122 insertions(+)
 create mode 100644 docs/experiment_round43_control_strat.json

diff --git a/docs/experiment_round43_control_strat.json b/docs/experiment_round43_control_strat.json
new file mode 100644
index 0000000..7f7ed35
--- /dev/null
+++ b/docs/experiment_round43_control_strat.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": 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_round43_control_strat.json",
+    "trace": null,
+    "select": "endpoint",
+    "replan": "empty",
+    "safety": 1,
+    "partial": 1,
+    "inc_cap": 1,
+    "gap_margin": 0.0,
+    "cert_batch": 0,
+    "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": 0.5234375,
+    "collision_free_success": 0.5234375,
+    "time_ok_1p3": 0.265625,
+    "collision_rate": 0.0,
+    "collision_steps_mean": 0.0,
+    "travel_ratio_median": 2.615795850753784,
+    "time_mean_s": 2.486567258834839,
+    "brakes_mean": 19.5546875,
+    "start_fail": 0.0,
+    "goal_ratio_2s": 0.27920001745224,
+    "err_fail_med": 204.71360778808594,
+    "err_success_med": 0.21751396358013153,
+    "dls_share": 0.61328125,
+    "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": 0.24741201102733612,
+    "reverse_mean": 0.25991496443748474,
+    "reverse_moving_p50": 0.32581019401550293,
+    "moving_share_p50": 0.9587156176567078,
+    "brake_rate_mean": 0.10028044879436493,
+    "brake_rate_p50": 0.0051282052882015705,
+    "brake_rate_p90": 0.3153846263885498,
+    "stuck_share": 0.234375,
+    "n_plans_mean": 195.0,
+    "success_reachable": 0.5514403581619263,
+    "by_cat": {
+      "A": [
+        128,
+        96
+      ],
+      "B": [
+        74,
+        20
+      ],
+      "C": [
+        32,
+        11
+      ],
+      "D": [
+        13,
+        0
+      ],
+      "Dr": [
+        9,
+        7
+      ]
+    }
+  },
+  "simulation_wall_s": 353.7248780000082
+}
diff --git a/docs/探索日志.md b/docs/探索日志.md
index 1056045..70d0524 100644
--- a/docs/探索日志.md
+++ b/docs/探索日志.md
@@ -108,6 +108,21 @@
 - **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。
 
+## 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`）：
+
+  | 新读数 | 值 | 独立离线预测 |
+  |---|---|---|
+  | `success_reachable`（排除按定义不可达的 D 类） | **0.5514** | 0.5514（逐题 dump 手算）✓ |
+  | `by_cat` 到达 | A 96/128=75%、B 20/74=27%、C 11/32=34%、D 0/13、Dr 7/9 | 与离线一致 ✓ |
+  | `brake_rate` mean / p50 / **p90** | 10.03% / 0.51% / **31.54%** | 均值与手算 19.55/195 一致 ✓ |
+  | `stuck_share`（brake_rate>0.2 的题占比） | **23.4%** | 与"最卡前 20% 题占 72.4% 刹停事件"同量级 ✓ |
+  | `reverse_p50` / `reverse_moving_p50` / `moving_share_p50` | 0.2474 / 0.3258 / 0.9587 | 离线预测 0.247 ✓ |
+
+- 三处独立来源（逐题 dump 手算、轨迹转储离线重算、评估内建指标）在四个数上完全对齐 ⇒ 这批分层读数可以当判据用，不只是看板。
+- **为什么现在必须有 p90/stuck_share**：对照口径的均值刹停率 10.03% 看着"差 5 倍"，但真相是 **p50 只有 0.51%、p90 高达 31.5%、23.4% 的题踩在卡住线上**。以前只有均值时，`backoff` 这类"专治尾巴症状"的改动能把判据刷好而行为不变（已实测：brake_rate −16%、到达 −0.8 pp）——**没有 p90/stuck_share 就看不出这种背离**，这就是 #25 存在的理由。
+
 ## `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`）：