From 671f1bdc89e553ee0fbcb4f5ec9b4fee7f9a5911 Mon Sep 17 00:00:00 2001
From: cbb <2625274797@qq.com>
Date: Tue, 20 Oct 2026 22:37:40 +0800
Subject: [PATCH] =?UTF-8?q?exp:=20=E8=BD=AE=E6=AC=A139=E2=80=94=E2=80=94K?=
 =?UTF-8?q?=3D16=20=E6=AD=A3=E5=90=91=E7=BB=93=E6=9E=9C=EF=BC=9A=E5=88=B0?=
 =?UTF-8?q?=E8=BE=BE+3.9pp/=E6=97=B6=E9=99=90+3.1pp/=E5=88=B9=E8=BD=A6-30%?=
 =?UTF-8?q?=E4=B8=94=E7=A2=B0=E6=92=9E=E4=BB=8D0=EF=BC=9B=E4=BB=AA?=
 =?UTF-8?q?=E8=A1=A8=E5=AF=B9=E8=A1=A8=201.575/6=3D0.2625=20vs=20=E7=A6=BB?=
 =?UTF-8?q?=E7=BA=BF=E9=A2=84=E6=B5=8B0.247?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 docs/experiment_round39_K16.json | 77 ++++++++++++++++++++++++++++++++
 docs/探索日志.md                 | 14 ++++++
 2 files changed, 91 insertions(+)
 create mode 100644 docs/experiment_round39_K16.json

diff --git a/docs/experiment_round39_K16.json b/docs/experiment_round39_K16.json
new file mode 100644
index 0000000..91509db
--- /dev/null
+++ b/docs/experiment_round39_K16.json
@@ -0,0 +1,77 @@
+{
+  "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_round39_K16.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": 0.5625,
+    "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": 1.5750000476837158,
+    "reverse_mean": 1.6161863803863525
+  },
+  "simulation_wall_s": 573.4911701000092
+}
diff --git a/docs/探索日志.md b/docs/探索日志.md
index 0794c48..44e798f 100644
--- a/docs/探索日志.md
+++ b/docs/探索日志.md
@@ -105,6 +105,20 @@
 - **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。
 
+## K16：候选数翻倍是本轮第一个正向结果（轮次 39 续）
+
+- 同一 256 题、同 seed、`--inc_cap 1`、其余配方不变，只把 `K` 从 8 加到 16（`docs/experiment_round39_K16.json`）：
+
+  | 口径 | 到达 | 无碰撞到达 | 时限内 | 刹车/题 | 行程比 |
+  |---|---|---|---|---|---|
+  | 对照 cap（K=8） | 0.5234 | 0.5234 | 0.2656 | 19.55 | 2.6158 |
+  | **K=16** | **0.5625** | **0.5625** | **0.2969** | **13.77** | 2.6243 |
+
+  ⇒ 到达 **+3.9 pp**、时限内 **+3.1 pp**、**刹车 −30%**，碰撞仍为 0，行程比不变。这一条直接支持失败分类的判断：**很多"全拒"不是真没路，而是 8 条候选里没有一条能走**；把候选池翻倍，安全层就找得到可行前缀。这是本轮（36–39）四个杠杆里第一个正向、且不放松任何判据的结果（限速 `inc_cap=1` 照开、间隙判据照旧）。
+- **代价没量之前不许上线**：K=16 意味着每次规划的候选点从 32 个（8×2×2）到 64 个，认证与打分都要翻倍——这正好是要用独占微基准量的东西（跨轮次墙钟不算证据：这批整批墙钟 573 s vs 对照 325 s，但同期我在同一台机器跑了 CPU 分析，**不能当成本**）。`steps_cost` 探针之后要补一个 `K=8 vs 16` 的 B=1 周期微基准。
+- **仪表对上了**：K16 报 `reverse_p50 = 1.5750`，是**我修分母之前**的口径（每步换向轴数，0–6），除以 6 = **0.2625**；离线按 rollout 同窗口算的对照臂预测是 **0.247** ⇒ 差 6%，在"K 更大→走得更顺"的合理方向内。读数规则写死：**产物里的 `reverse_p50` 早于 14:28 的一律先除 6**，之后的才是 [0,1] 的分数。
+- 之后三臂（`safety_try 3/1`、`z_scale 1.5`）在跑：`safety_try` 是**成本方向**的对照（只试前 3/1 条候选会掉多少到达），`z_scale` 是多样性方向。
+
 ## 指标链路端到端自证 + 预注册一条判读（轮次 39 续）
 
 - **CPU 冒烟（4 题、`--select ep_waste --len_w 0.3 --cert_batch 1 --backoff 2 --inc_cap 1`）**：产物里 `reverse_p50 / reverse_mean / backoff_mean / reject_all_mean` 四个新键**都在**（`ml/runs/_smoke_waste.json`）。这是在验证"生产端改了、消费端也要改"那条教训落地——上次 `backoff_mean` 就是这样静默丢掉的。同时证明 `ep_waste` 与 `backoff` 在真实运动学管线上能一起跑，不靠打桩。