From 422fb8b72f4e897fd7aa3f80f8b807f35ecd737d Mon Sep 17 00:00:00 2001
From: cbb <2625274797@qq.com>
Date: Tue, 20 Oct 2026 22:10:44 +0800
Subject: [PATCH] =?UTF-8?q?exp+fix:=20=E8=BD=AE=E6=AC=A139=E2=80=94?=
 =?UTF-8?q?=E2=80=94backoff=3D0+cert=5Fbatch1=20=E4=B8=8E=E8=BD=AE?=
 =?UTF-8?q?=E6=AC=A133=20cap=20=E9=80=90=E9=A1=B9=E7=9B=B8=E5=90=8C(?=
 =?UTF-8?q?=E9=87=8D=E5=86=99=E6=89=B9=E9=87=8F=E5=88=86=E6=94=AF=E6=97=A0?=
 =?UTF-8?q?=E8=A1=8C=E4=B8=BA=E5=8F=98=E5=8C=96)=EF=BC=9B=E4=BF=AE=20backo?=
 =?UTF-8?q?ff=20=E6=8C=87=E6=A0=87=E8=A2=AB=20eval=5Fcl=20=E6=98=BE?=
 =?UTF-8?q?=E5=BC=8F=E9=94=AE=E5=88=97=E8=A1=A8=E4=B8=A2=E6=8E=89?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 docs/experiment_round38_backoff0.json | 73 +++++++++++++++++++++++++++
 docs/探索日志.md                      |  4 ++
 ml/eval/eval_cl.py                    |  3 ++
 3 files changed, 80 insertions(+)
 create mode 100644 docs/experiment_round38_backoff0.json

diff --git a/docs/experiment_round38_backoff0.json b/docs/experiment_round38_backoff0.json
new file mode 100644
index 0000000..c3945c8
--- /dev/null
+++ b/docs/experiment_round38_backoff0.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": null,
+    "out": "docs/experiment_round38_backoff0.json",
+    "trace": null,
+    "select": "endpoint",
+    "replan": "empty",
+    "safety": 1,
+    "partial": 1,
+    "inc_cap": 1,
+    "gap_margin": 0.0,
+    "cert_batch": 1,
+    "steps": 1,
+    "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
+  },
+  "simulation_wall_s": 329.1139315000037
+}
diff --git a/docs/探索日志.md b/docs/探索日志.md
index 5303af8..30637c9 100644
--- a/docs/探索日志.md
+++ b/docs/探索日志.md
@@ -47,6 +47,10 @@
   - 待测：`len_w ∈ {0.1, 0.3, 1.0}` 三臂，看行程比能压到多少、到达与时限各付多少。这条杠杆与"到达率"是相互冲突的目标（惩罚行程=主动走慢），所以必须成对汇报，不能只报行程变好。
   - 单位口径要说清：`len_w` 的量纲是"mm 误差 / 一个归一 STEP 的累计行程"，16 步满速在 3 条轴上大约是 48，所以 0.1~1.0 对应的惩罚是 5~50 mm 当量——这是扫这三个值的原因，不是随手取的。
 - 测试 `tests/test_eval_cl.py` 38→40 项、全套 43→45 项（`python -m unittest discover -s tests`，2.2 s 全绿）。GPU 队列顺序：backoff A/B → `--safety 0` 判别 → 候选档位 → 行程惩罚。
+- **等价闸门先过了**：`--backoff 0 --cert_batch 1 --inc_cap 1` 的 256 题与轮次 33 cap（`--cert_batch 0`）**逐项完全相同**（`docs/experiment_round38_backoff0.json` 对 `docs/experiment_round33_cap.json`，把两个 metrics 字典做键并集比较，差异集合为空，含 success 0.523438、time_ok 0.265625、brakes 19.55…）⇒ 轮次 36 我把批量认证分支重写（加缩步档那一段）没有改动任何行为。
+- **两个我自己的操作失误，如实记在这里**：
+  1. 新加的 `backoff` 指标一度在产物里根本没有——`ml/eval/eval_cl.py:run()` 的 `out` 是**自己显式列键**的，不是从 `rollout` 返回的字典通抄，所以我往 `res` 加的 `backoff_mean` 被无声丢掉。已改成从逐题 `per['backoff']` 汇总（`backoff_mean` + `reject_all_mean = backoff+brakes`，后者才是与"原来全拒次数"可比的量）。**教训：加指标要顺着输出链路走到底，别只改生产端。**
+  2. 停后台 GPU 作业时，`TaskStop` 之后 `bash` 脚本链会活下来并继续拉起下一个 python；而 `nvidia-smi --query-compute-apps` 在这台机器上列不出这些 python（只列 ShellHost/wallpaper32 之类），据此判断"卡已空"会误判。我第一次按 PID 杀子进程后，链条 30 秒后又起了一个新的 eval。第二次我按进程树杀，但**把父子关系读反了**（Windows 的 PPID 在父进程已死时是陈旧值），结果误杀了新链的第一臂。正确做法：重启前用命令行正则列出全量候选、数 python 个数确认单租户，重启后再数一次；杀之前先按 `StartTime` 判新旧。backoff 那一臂因此要补跑。
 
 ## 轮次 38：把 `--steps` 打通之后，闭环反而变差——瓶颈从"动作质量"移到了"安全层拒绝率"
 
diff --git a/ml/eval/eval_cl.py b/ml/eval/eval_cl.py
index a5823d7..bfcd66d 100644
--- a/ml/eval/eval_cl.py
+++ b/ml/eval/eval_cl.py
@@ -49,6 +49,9 @@ def run(model, kin, tasks, device, bs=256, **kw):
         out.update({'speed_ratio_p50': median(v), 'speed_ratio_p95': v.quantile(0.95).item(),
                     'speed_ratio_max': v.max().item(), 'over_vmax_rate': (v > 1.001).float().mean().item(),
                     'accel_ratio_p95': acc.quantile(0.95).item(), 'accel_ratio_max': acc.max().item()})
+    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()
     return out, per