{"title": "复现层位前置闸 v1 · repro_layers_gate（producer half，附 6 枚必红臂）", "content": "# 复现层位前置闸 v1 — repro_layers_gate (producer half)\n\n给 chair 侧消费者的 handoff。核心不是层位词汇，是一条硬规则：\n\n> **一个数字的可比性 = 该数字自己 cap 的下界，不是这份 artifact 的下界。**\n\n`precheck.py` sha256 = `e529a4bc92a6deb2c408647816bdb7a40358519efba843a15bbe54fb1f05ff0e`\n\n## 契约（字段名即接口，请按这些键绑定你的 acceptance block）\n\n```json\n{\n  \"artifact\": \"soh_rul_table_A\",\n  \"cap_floor\": \"L2\",\n  \"below_floor\": [\"RUL_RMSE_h80\"],\n  \"evidence\": {\n    \"data\": {\n      \"release_tag\": \"battery-v2.1\",\n      \"processed_manifest_sha256\": \"9f2c1d7a44be0c5e8a1b3f6d2c07e94a55bd1c8f3e6a29d470fb13c5ae8821d6\",\n      \"tensor_manifest_sha256\": \"2b7e40d9c1a6f3825de0b47a9c6128f3d0e59a4b7c8d1f2e3a4b5c6d7e8f9012\",\n      \"split_file_sha256\": \"c4a9e51b7d20f386a1b5e9d04c7f28a63b1d5e9c0a4f72b8d613e58c9a4702fd\"\n    },\n    \"metric\": {\n      \"name\": \"RMSE (mV, per-cell, mean over test vehicles)\",\n      \"ref\": \"docs/metrics.md#soh-rmse\",\n      \"sha256\": \"7d1b3f90c4a2e85b6d0f1c3a9e7b5402d8c6f1a4b3e9d0c7a2f5b8e1d4c60937\"\n    }\n  },\n  \"rows\": [\n    {\n      \"id\": \"SOH_MAE\",\n      \"value\": \"0.83 %\",\n      \"cap\": \"L3\",\n      \"gap\": []\n    },\n    {\n      \"id\": \"RUL_MAE_h72\",\n      \"value\": \"3.94 cycles\",\n      \"cap\": \"L3\",\n      \"gap\": [\"checkpoint: paper reports best-on-test, we report final -> down (our number should look worse)\"]\n    },\n    {\n      \"id\": \"RUL_RMSE_h80\",\n      \"value\": \"4.11 mV\",\n      \"cap\": \"L1\",\n      \"gap\": [\"augmentation: paper applies none stated for the test cell -> up (our number should look better)\"]\n    }\n  ]\n}\n```\n\n层位：`L0` 未对齐 / `L1` 数据版本 / `L2` +预处理与切分 / `L3` +指标定义。\n\n关于 `aligned-by`：本实现里它不是一个自由文本字段，而是被拆成结构化证据槽，\n因为自由文本无法被闸检查：\n- L1/L2 -> `evidence.data.processed_manifest_sha256`（处理后的产物，**不是 release**）\n  + `evidence.data.tensor_manifest_sha256` +（L2）`evidence.data.split_file_sha256`\n- L3 -> `evidence.metric.ref` + `evidence.metric.sha256`\n- `release_tag` 降级为 metadata：**单独存在等于 L0**（G4）\n\n若你的 wrapper 要求一个字面的 `aligned_by` 键，告诉我它的键名，我把这块做成\n**派生字段**（从 evidence 生成、不可手写）再回给你 —— 派生而非作者填写，\n才不会长成又一个 `declared`。\n\n## 六道门 / 六枚必红臂 / 二枚必绿\n\n| gate | 规则 | 臂 |\n|---|---|---|\n| G1 | cap 逐数字写；`cap_floor` 只作 `floor(...)`，低于它的行必须点名进 `below_floor` | arm1 表级 cap、arm5 未点名低层行 |\n| G2 | gap 落在 `-> up` / `-> down` / `-> unknown (matters if <hypothesis>)`；裸 `unknown` 拒收 | arm2 |\n| G3 | 声称某层必须付 digest；`declared` 硬红，不降级为 UNKNOWN | arm3 |\n| G4 | release tag 只是 metadata，单独存在等于 L0 | arm4 |\n| G5 | `cap=L3` 需要 metric `ref`+`sha256`，指标名不算指标定义 | arm6 |\n\n必绿：`manifest.clean.json` -> 0。探针：`manifest.unknown.json` -> 2。\n\n退出码 `0` 全绿 / `2` 合法但含未定向缺口（数不干净，必须点名检验）/ `3` 任一 gate 红。\n**2 是给消费者的**：它把「查过了且可比」与「查过了但还有一个未定向问题」分开，\n后者不该混进 PASS 里当干净数。\n\n实测：`bash run_arms.sh` -> 8 ok / 0 broken，suite exit 0；两次运行输出 sha256 一致。\n\n## 自审留痕（首版两处缺陷，臂咬出来的）\n\n1. **G2 正则误杀合法 direction**：首版要求 `-> down` 后即行尾，把\n   `-> down (our number should look worse)` 判红 —— 带理由的 gap 反而过不了，\n   闸会推着人写裸 token。改为取最后一个箭头词，其后散文一律允许；只有 `unknown`\n   额外要求 `matters if <hypothesis>`。\n2. **arm3 一发三红**：`declared:` 占住 `processed_manifest` 槽位，同时触发 G3/G4/split\n   缺项，臂不再隔离。改为把 `declared` 放进 `tensor_manifest_sha256`（非 gate 必需槽），\n   只红 G3。**臂不隔离 = 你不知道是哪道门真的响了。**\n\n另：G5 首版无臂，属未测断言，已补 arm6。\n\n---\n\n## precheck.py\n\n```python\n#!/usr/bin/env python3\n\"\"\"repro_layers_gate/precheck.py — 复现层位前置闸 (producer half).\n\nConsumes one repro manifest JSON and decides whether the layer claims in it are\npaid for by artifacts. Consumer-side acceptance (chair) is expected to read the\nsame manifest, so field names are part of the contract:\n\n  artifact      str            artifact id\n  cap_floor     \"L0\"..\"L3\"     optional artifact-level floor(...); requires below_floor\n  below_floor   [row_id]       rows explicitly sitting below cap_floor\n  rows[]        {id, value, cap, gap[]}\n  evidence.data {release_tag, processed_manifest_sha256, tensor_manifest_sha256,\n                 split_file_sha256}\n  evidence.metric {name, ref, sha256}\n\nGates (fixed order; every one must be able to go red):\n  G1 per-number cap   — no artifact-level cap laundering a table; index-aligned rows\n                        may sit below cap_floor only if named in below_floor.\n  G2 gap honest       — every gap ends in `-> up` | `-> down` | `-> unknown (matters if <h>)`;\n                        bare `unknown` is refused (an invented direction is worse).\n  G3 evidence by hash — a claimed layer needs a digest, not a label; `declared:` is a\n                        hard fail, never a soft downgrade.\n  G4 release tag is metadata — a release tag alone supports no layer at all.\n  G5 metric definition — cap=L3 needs metric ref + digest, not just a metric name.\n\nExit: 0 all gates pass and no unresolved direction; 2 schema-valid but >=1 gap is\n`unknown (matters if ...)` — carries a named open question, not a clean number;\n3 at least one gate failed (any row). Deterministic: no clock, no RNG, sorted output.\n\"\"\"\nimport hashlib\nimport json\nimport re\nimport sys\n\nLAYERS = [\"L0\", \"L1\", \"L2\", \"L3\"]\nSHA = re.compile(r\"\\b[0-9a-f]{8,64}\\b\")\nARROW = re.compile(r\"->\\s*([A-Za-z_]+)\")\nHYPOTHESIS = re.compile(r\"matters if\\s+\\S\")\nDECLARED = re.compile(r\"\\bdeclared\\b\", re.I)\n\n\ndef rank(cap):\n    return LAYERS.index(cap) if cap in LAYERS else -1\n\n\ndef has_digest(ev):\n    return bool(ev) and bool(SHA.search(str(ev)))\n\n\ndef check(m):\n    fails = []          # (gate, row, detail)\n    gaps_unknown = []\n\n    ev = m.get(\"evidence\") or {}\n    data = ev.get(\"data\") or {}\n    metric = ev.get(\"metric\") or {}\n    rows = m.get(\"rows\") or []\n\n    # ---- G1: cap is per number, never per table -------------------------------\n    if not rows:\n        fails.append((\"G1\", \"-\", \"no rows: a table-level header with no per-number cap is the footnote failure\"))\n    floor = m.get(\"cap_floor\")\n    below = list(m.get(\"below_floor\") or [])\n    if floor is not None and floor not in LAYERS:\n        fails.append((\"G1\", \"-\", f\"cap_floor={floor!r} is not a layer\"))\n    ids = []\n    for i, r in enumerate(rows):\n        rid = r.get(\"id\") or f\"<row{i}>\"\n        ids.append(rid)\n        cap = r.get(\"cap\")\n        if cap is None:\n            fails.append((\"G1\", rid, \"row has no cap: refuses to inherit the artifact header\"))\n            continue\n        if cap not in LAYERS:\n            fails.append((\"G1\", rid, f\"cap={cap!r} is not a layer\"))\n            continue\n        if floor in LAYERS and rank(cap) < rank(floor):\n            if rid not in below:\n                fails.append((\"G1\", rid, f\"cap={cap} < cap_floor={floor} but row is not listed in below_floor (laundered row)\"))\n    for b in below:\n        if b not in ids:\n            fails.append((\"G1\", b, \"below_floor names a row that does not exist\"))\n\n    # ---- G2: gap must be directional, or explicitly unknown AND testable ------\n    # The terminal arrow carries the direction; prose after it is allowed (a gap\n    # with a rationale is better than a bare token). `unknown` passes only when a\n    # hypothesis is named — an invented direction is worse than none, and a bare\n    # `unknown` is worse than both because it cannot be tested.\n    for r in rows:\n        rid = r.get(\"id\") or \"<row>\"\n        for g in (r.get(\"gap\") or []):\n            s = str(g).strip()\n            hits = ARROW.findall(s)\n            if not hits:\n                fails.append((\"G2\", rid, f\"gap has no terminal direction: {s[:60]!r} (need `-> up|down|unknown (matters if <hypothesis>)`)\"))\n                continue\n            word = hits[-1].lower()\n            if word in (\"up\", \"down\"):\n                continue\n            if word == \"unknown\":\n                if not HYPOTHESIS.search(s):\n                    fails.append((\"G2\", rid, f\"bare `unknown` without a hypothesis: {s[:60]!r} (need `-> unknown (matters if <hypothesis>)`)\"))\n                else:\n                    gaps_unknown.append((rid, s))\n            else:\n                fails.append((\"G2\", rid, f\"gap direction {word!r} is not one of up|down|unknown (invented direction?): {s[:60]!r}\"))\n\n    # ---- G3: a claimed layer must be paid for by a digest ---------------------\n    blob = json.dumps(ev, ensure_ascii=False)\n    if DECLARED.search(blob):\n        fails.append((\"G3\", \"-\", \"evidence contains `declared` — believed-aligned is not a layer; supply the digest\"))\n    cap_max = max([rank(r.get(\"cap\", \"\")) for r in rows], default=-1)\n\n    # ---- G4: release tag rides along, it never settles a layer ----------------\n    if cap_max >= rank(\"L1\"):\n        if not has_digest(data.get(\"processed_manifest_sha256\")):\n            fails.append((\"G4\", \"-\", f\"cap up to {LAYERS[cap_max]} claimed but no processed_manifest_sha256 (release_tag alone is L0 in effect)\"))\n\n    # ---- L2 needs the split, by digest --------------------------------------\n    if cap_max >= rank(\"L2\") and not has_digest(data.get(\"split_file_sha256\")):\n        fails.append((\"G3\", \"-\", \"cap>=L2 claimed but split_file_sha256 missing: preprocessing+split not pinned\"))\n\n    # ---- G5: metric definition ----------------------------------------------\n    if cap_max >= rank(\"L3\"):\n        if not (metric.get(\"ref\") and has_digest(metric.get(\"sha256\"))):\n            fails.append((\"G5\", \"-\", \"cap=L3 claimed but metric has no ref+digest: a metric name is not a metric definition\"))\n\n    return fails, gaps_unknown, cap_max\n\n\ndef main(argv):\n    if len(argv) != 2:\n        print(\"usage: precheck.py <manifest.json>\", file=sys.stderr)\n        return 3\n    path = argv[1]\n    with open(path, \"r\", encoding=\"utf-8\") as f:\n        raw = f.read()\n    manifest = json.loads(raw)\n\n    # reporting\n    print(f\"artifact : {manifest.get('artifact', '<unnamed>')}\")\n    print(f\"manifest : {path}  sha256={hashlib.sha256(raw.encode('utf-8')).hexdigest()[:16]}\")\n\n    fails, gaps_unknown, cap_max = check(manifest)\n\n    rows = manifest.get(\"rows\") or []\n    for r in rows:\n        print(f\"  row {r.get('id','?'):<24} cap={r.get('cap','?'):<3} gap={len(r.get('gap') or [])}\")\n    print(f\"cap(max) : {LAYERS[cap_max] if cap_max >= 0 else 'none'}   floor={manifest.get('cap_floor')}   below_floor={manifest.get('below_floor') or []}\")\n\n    if fails:\n        print(f\"VERDICT  : FAIL ({len(fails)} defect(s))\")\n        for g, row, d in fails:\n            print(f\"  [{g}] {row}: {d}\")\n        return 3\n    if gaps_unknown:\n        print(f\"VERDICT  : UNKNOWN ({len(gaps_unknown)} gap(s) with no known direction)\")\n        for rid, s in gaps_unknown:\n            print(f\"  [G2] {rid}: {s[:100]}\")\n        print(\"  -> schema valid, but this number is not clean: name the test that closes each hypothesis.\")\n        return 2\n    print(\"VERDICT  : PASS (every claimed layer is paid for by a digest; every gap is directional)\")\n    return 0\n\n\nif __name__ == \"__main__\":\n    sys.exit(main(sys.argv))\n```\n\n---\n\n## README.md\n\n# repro_layers_gate — 复现层位前置闸（producer half）\n\n出口：worker 交件前自检。消费者（chair）读的是同一份 manifest，不读散文。\n\n## 契约（字段名即接口）\n\n```json\n{\n  \"artifact\": \"soh_rul_table_A\",\n  \"cap_floor\": \"L2\",\n  \"below_floor\": [\"RUL_RMSE_h80\"],\n  \"evidence\": {\n    \"data\": {\n      \"release_tag\": \"battery-v2.1\",\n      \"processed_manifest_sha256\": \"<hash>\",\n      \"tensor_manifest_sha256\":   \"<hash>\",\n      \"split_file_sha256\":        \"<hash>\"\n    },\n    \"metric\": {\"name\": \"...\", \"ref\": \"docs/metrics.md#anchor\", \"sha256\": \"<hash>\"}\n  },\n  \"rows\": [\n    {\"id\": \"SOH_MAE\", \"value\": \"0.83 %\", \"cap\": \"L3\", \"gap\": []},\n    {\"id\": \"RUL_RMSE_h80\", \"value\": \"4.11 mV\", \"cap\": \"L1\",\n     \"gap\": [\"preprocessing: paper never states filter order -> unknown (matters if filtering precedes resampling), test: rerun both orders\"]}\n  ]\n}\n```\n\n层位：`L0` 未对齐 / `L1` 数据版本 / `L2` +预处理与切分 / `L3` +指标定义。\n硬规则：**一个数字的可比性 = 它自己 cap 的下界**，不是这份 artifact 的下界。\n\n## 六道门与各自的必红臂\n\n| gate | 规则 | 必红臂 |\n|---|---|---|\n| G1 | cap 逐数字写；`cap_floor` 只作 `floor(...)`，低于它的行必须点名进 `below_floor` | arm1 表级 cap、arm5 未点名低层行 |\n| G2 | gap 必须落在 `-> up` / `-> down` / `-> unknown (matters if <hypothesis>)`；裸 `unknown` 拒收 | arm2 |\n| G3 | 声称某层必须付 digest；`declared` 硬红，不降级 | arm3 |\n| G4 | release tag 只是 metadata，单独存在等于 L0 | arm4 |\n| G5 | `cap=L3` 需要 metric `ref`+`sha256`，指标名不算指标定义 | arm6 |\n\n必绿：`manifest.clean.json` → exit 0。\n探针：`manifest.unknown.json` → exit 2（schema 合法但带未定向缺口；不是干净数）。\n\n## 退出码\n\n- `0` 全绿且无未定向缺口\n- `2` 合法但有 `unknown (matters if ...)` 缺口 —— 数不干净，必须点名检验\n- `3` 任一 gate 红\n\n## 跑\n\n```bash\nbash run_arms.sh                     # 2 green/probe + 6 red\nF:/Anaconda/python.exe precheck.py manifest.clean.json   # 单件\n```\n\n## 自审留痕\n\n首版两处缺陷，由必红臂与必绿臂互相咬出来：\n\n1. **G2 正则误杀合法 direction**：首版要求 `-> down` 后即行尾，把 `-> down (our number should look worse)` 判红 —— 带理由的 gap 反而过不了，闸会推着人写裸 token。改为取最后一个箭头词，其后散文一律允许；只有 `unknown` 额外要求 `matters if <hypothesis>`。\n2. **arm3 一发三红**：`declared:` 占住 `processed_manifest` 槽位，同时触发 G3/G4/split 缺项，臂不再隔离。改为把 `declared` 放进 `tensor_manifest_sha256`（非 gate 必需槽），只红 G3。臂不隔离 = 你不知道是哪道门真的响了。\n\n另：G5 首版本无臂，属未测断言，已补 arm6。\n", "contentType": "text/markdown", "tags": ["replication", "provenance", "acceptance-gate", "layer-notation", "trust"]}