From 27c8cbf1119716a81ff37d62c59f4ba4be08a3ad Mon Sep 17 00:00:00 2001
From: cbb <2625274797@qq.com>
Date: Tue, 20 Oct 2026 23:03:06 +0800
Subject: [PATCH] =?UTF-8?q?exp:=20=E8=BD=AE=E6=AC=A141=E2=80=94=E2=80=94K1?=
 =?UTF-8?q?6=E4=B8=B2=E8=A1=8C/=E6=89=B9=E9=87=8F=E5=86=B3=E7=AD=96?=
 =?UTF-8?q?=E9=80=90=E4=BD=8D=E7=9B=B8=E5=90=8C(=E7=AD=89=E4=BB=B7?=
 =?UTF-8?q?=E6=80=A7=E6=89=A9=E5=88=B0K=3D16)+=E5=A2=99=E9=92=9F-18%?=
 =?UTF-8?q?=EF=BC=9Bbrake=5Frate=E9=A6=96=E8=AF=BBmean7.06%=E4=BD=86p50=3D?=
 =?UTF-8?q?0=EF=BC=8C=E5=88=B9=E5=81=9C=E9=9B=86=E4=B8=AD=E5=9C=A8?=
 =?UTF-8?q?=E4=B8=80=E5=B0=8F=E6=92=AE=E5=8D=A1=E4=BD=8F=E7=9A=84=E9=A2=98?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 docs/experiment_round40_K16cb1.json | 82 +++++++++++++++++++++++++++++
 docs/探索日志.md                    | 10 ++++
 2 files changed, 92 insertions(+)
 create mode 100644 docs/experiment_round40_K16cb1.json

diff --git a/docs/experiment_round40_K16cb1.json b/docs/experiment_round40_K16cb1.json
new file mode 100644
index 0000000..62201ee
--- /dev/null
+++ b/docs/experiment_round40_K16cb1.json
@@ -0,0 +1,82 @@
+{
+  "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": null,
+    "out": "docs/experiment_round40_K16cb1.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,
+    "n_plans_mean": 195.0
+  },
+  "simulation_wall_s": 469.2955887999997
+}
diff --git a/docs/探索日志.md b/docs/探索日志.md
index a2e0a2b..c73d63b 100644
--- a/docs/探索日志.md
+++ b/docs/探索日志.md
@@ -108,6 +108,16 @@
 - **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。
 
+## K=16 串行 vs 批量认证等价 + 第一次真正核算"刹停率"（轮次 41）
+
+- `--K 16 --cert_batch 1` 与 `--K 16`（串行）跑同一 256 题同 seed：`success / time_ok_1p3 / brakes_mean / travel_ratio_median / collision_*` **全部逐位相同**，差异键只有 7 个、且全是我这几轮新加/改口径的报表键（`reverse_*` 的两种口径、`n_plans`、`brake_rate`、`moving_share`）。⇒ **轮次 39 的"试候选数 1/3/8/16 曲线"同样描述批量口径**，也就是说 K=16 到达 0.5625 这个收益不依赖串行早停这个巧合。
+  （`docs/experiment_round40_K16cb1.json` vs `docs/experiment_round39_K16.json`）
+- **整批墙钟 573 s → 469 s（−18%）**：批量把 K 次小调用并成 1 次大调用，在 B=256 的评估里本来就更快。这只说明趋势，**单次规划的 P99 要用独占微基准另测**（修好的探针排在链尾）。
+- **`brake_rate`（刹停 / 规划次数）第一次有读数**：K=16 口径 `mean = 7.06%`，但 **`p50 = 0.0`**，`n_plans_mean = 195`，`moving_share_p50 = 98.7%`。
+  ⇒ §9.2 要"<2% 推理周期"，均值差 3.5 倍，但**中位任务一次都不刹**——全部刹停集中在一小撮题上。这与失败分类完全一致（63/109 是"走一半被冻住"，而它们就是那撮尾巴）。
+  **含义**：继续在"普遍少刹"上花力气收益有限，真正的问题是**那批卡住的题里安全层没有可行候选**——正是 K=16（+3.9 pp）与 #24（把间隙喂进打分）要打的靶子。
+- 仪表互验：K=16 批量臂 `reverse_p50 = 0.2625` 与我离线同窗口重算的 K=8 对照 0.247 同量级（K=16 略高，与它到达更多、动得更多一致）；`reverse_moving_p50 = 0.3194` > 全步口径 ✓ 方向正确（除掉站住的步之后更抖）。之前那批 1.x 的读数确系修分母前的旧口径。
+
 ## #24 的设计定稿：一次几何调用同时供打分与认证（不放松判据、不加成本）
 
 - 表面看有循环依赖：打分要用的"沿路间隙"目前只在安全层里算，而安全层只认证**排完序**的候选。破法是把顺序倒过来——**先对全部 K 条候选算一次 `2·commit_n` 点的间隙矩阵**，然后：