Download docs/BATCH_LEDGER.md from weideyouni/holoaeon: direct link, hf CLI and curl.
- Browser
- Download file 81.4 kB
-
https://huggingface.co/weideyouni/holoaeon/resolve/main/docs/BATCH_LEDGER.md
- Command line
-
hf download hf://weideyouni/holoaeon/docs/BATCH_LEDGER.md
-
curl -L -o BATCH_LEDGER.md https://huggingface.co/weideyouni/holoaeon/resolve/main/docs/BATCH_LEDGER.md
BATCH_LEDGER(历史批次台账 · append-only)
本文件由 V88-P2B 瘦身从 README.zh-CN.md / README.md 移入;历史批次行逐字保留、禁改写、禁删史。 活体台账(最近 3 批)仍在两份 README 的「当前工程状态」表;本文件只承载历史行。 建立:2026-11-19 · 桃花庵出品 · 封存密令「潮起开门,潮落闭嘴」。
中文历史台账(移出 README.zh-CN.md)
| 批次 | 内容 |
|---|---|
| 上一批次 | V84 恒真判据常驻金哨(tautology_scan A 口径 + 白名单登记账 + v3_gate 第 34 面 A14) —— V82/V83 只治了两处已知恒真,下一处仍要靠人肉发现;本批把「扫描」变成常驻闸:①新模块 holoaeon/tautology_scan.py(纯 AST、零执行、零新依赖、py3.8 兼容)静态检测「两常量比较」四类形态(tautology_eq / dead_always_true / dead_always_false / membership_in_const);②data/tautology_allowlist.json 登记 8 处正当用途(常量不变式自检 / 配置常量分派 / dict 声明在位检查),每条带 liveness 活性来源——白名单是登记账不是放行名单(缺 liveness / 条目过期 / 摘要不符一律红);③scripts/v3_gate.py 长出第 34 面 A14,三判据 unregistered_tautology · allowlist_drift · scanner_selfcheck;④扫描器自检吸收双方 v1 教训(四人组 25 处链式比较误报、执行端 16 处常量环境误报);⑤8 金丝雀 C1–C8。静态口径只能判 B 口径(字面常量比较):未登记命中 = 待复核,不等于已判定为病;A 口径(判据是否真约束输入)仍需人工读上下文。 ci.yml 字节不变(连续第 6 批)· DOC 0.9.77 → 0.9.78 |
| 上一批次 | V83 同型恒真第二处处置(topology_resonance.quartet_tensor 的 P2×P3 真实入参) —— V82 治了 quartet_director 的恒真判据,并把「同型恒真第二处」写进 honest_boundaries;本批治它:①quartet_tensor 增 p23 真实入参,**缺省直读本函数自有 _cosine(vecs["P2"], vecs["P3"])**(vecs = [s, 1-s] 二维表示,该函数本来就靠它算 res_pairs)——这是真家伙不是约定,与 V82 的 ``1- |
| 上一批次 | V82 P1-B 失效判据处置(两条恒真/伪影判据从披露走到处置) —— V81 边界只「披露」不「处置」,本批把两条源码级失效判据真正改掉:①advisor_gradient_flow 的 governable 从布尔压三态改为 True / False / None,曲率判定加相对容差 ``tol = 1e-9 · max(1.0, |
| 上一批次 | V80 云端环境差异可失败化(ENV_DELTA 注册簿) —— 把「本地全绿、云端少跑 46 个测试却同样显示绿」从口径运气变成机器可判据:claims_gate 长第 11 面 env_delta_consistency + data/env_delta_registry.json(D01–D11),三条可失败判据 unregistered_guard(每处环境守卫必须归属某条目)/ parity_disclosed(双端 ran·skipped·executed 差值按套逐条可解释、README 双口径同屏回读)/ digest_drift(行锚 ±10 内须有探测因子、自摘要改一字节且不重算即红),八只金丝雀逐只断言判红;先证红(未登记态 25/25 守卫全红、mismatch 29)再转绿。云端 py3.12(V80 推送后 run 43540948562)Ran 2031 / skipped 33、契约 184 / 11 与本机 2036 / 3、184 / 0 的差值全部有名有姓(F 盘素材 30 skip + 5 未发现、客仓 10、交付包 1、仓外模块 3 同态);已治项(CRLF / 时区)只作谱系、不充数解释现行差值。零新增模块、ci.yml 字节不变(仍在第 9 道 verify 内扩容);面数一律 python -m holoaeon.claims_gate report 回读、不写死 | V79 跨账口径一致性闸(claims_gate 第 10 面,专治「把技术阻塞误记成碳基意图」) —— 第 9 面已被 V81 regulation_semantic_consistency 占用,本批顺延为第 10 面 cross_ledger_consistency:注册簿新增 ledger_sources 七权威源(carbon_intent 口述终裁 高于一切派生账本,冲突时判记录红、不判碳基红)与 ledger_conflicts D0–D5;四条可失败判据 unregistered_conflict / stale_pointer / intent_vs_blocker / drift_after_register,八只金丝雀逐只断言判红。零新增模块、ci.yml 字节不变(复用已在跑的第 9 道 python -m holoaeon.claims_gate verify);面序号一律以 python -m holoaeon.claims_gate report 回读为准,不写死 |
| 上一批次 | V78b CI 门禁正式收口(复算闸与 v3 域闸入 ci.yml,10 → 12 步) —— P0-A 行尾归一化:check_sums 对文本文件按 LF 字节口径 哈希(二进制按 .gitattributes 的 binary 行原字节,规则唯一来源,不在闸里硬编码),Windows CRLF 工作区与 Linux checkout 复算同一份清单——「本地绿=云端绿」从口号变成判据;P0-B 清单卫生:移出 5 条运行期产物(reports/daily_baseline_20261103.md 等 3 条未跟踪 + 2 条 gitignore 命中,云端 checkout 上不存在、入清单必红),并给闸新增 untracked / ignored 两条机器判据——运行期产物入清单从此被拦,不再靠人肉对账;重建器 build_b8_dual_sync.release_files 同步只收 git 已跟踪文件;P0-C Manifest integrity(复算闸)与 v3 gate(33 面)两个步骤正式进 ci.yml,ci_preflight 契约同步 10 → 12 面(A13 双向对齐随之扩容,test_v78_ci_preflight 断言同批改);claims_gate._receipt 改按回执自身 doc_version 取最新(字母后缀批次 V78b 晚于纯数字批次 V81,按文件名 N 排序会指错回执——「最新」的真判据是 doc_version)。重建 SHA256SUMS:251 → 332 条(移 5 + 追平 V 系批次新增的 86 条已跟踪件)。DOC 0.9.72 → 0.9.73(四处耦合同批);本地全绿:主套 1931 OK · v3_gate 33/33 · claims 九面 · 复算 332 条 0 不匹配。 |
| 上一批次 | V81 规定契约语义一致性 + 通用新鲜度契约(claims_gate 长出一面,专治「登记了,但运行时不这样」) —— 旧账不是「README 少写一个版本号」:README.md 写 MCP 单版本 2024-11-05、mcp_adapter.py 的 PROTOCOL_VERSION 默认值也是它、注册簿却把「当前口径」登记成 2026-07-28 —— 三处各说各话,而所有闸全绿。不新建模块;**ci.yml 字节不变(sha256[:16] 前后同值 b743704ca8c307f7)。新增一面 regulation_semantic_consistency(面序号以 python -m holoaeon.claims_gate report 回读为准),三条判据:①每个 claim_site 逐条回读,出现未登记版本 token 即 claim_site_drift;②runtime_pointer + runtime_role/runtime_value 把「运行时真正生效的常量」登记进册,legacy 必须带 superseded_by;③freshness_contract 用一套** schema 同时管外部针与内部实测基线 —— risk_tier 分级 fast 90 / annual 180 / stable 365,source_sha16 一漂移即红,change_probe.mode 只认 manual_review_only(离线项目不声称已联网核实)。注册簿:6 条 pin 补六字段 + 4 条内部基线,measured_at 与文件 mtime 逐条对齐,未倒填。先证红再转绿:未扩字段时实测 42 处红 / pins_with_contract=0;8 只金丝雀 8/8 全红,C1、C2 分别打在真实 R1(README.md#L364 单版本)与真实 R2(runtime_pointer 缺失);补登记后 PASS / mismatch 0。新增 test_v81_regulation_semantic.py 53 例(十组);主套 1874 → 1931;redline 156;DOC 0.9.71 → 0.9.72。边界:本批只让脱钩可失败,**不擅自把 PROTOCOL_VERSION 改成 2026-07-28**(兼容性归碳基);verified_at 仍是人写快照,闸只判漂移与过期,不判规定是否现行有效。 |
| 上一批次 | V78a 本地预检闸(把「本地跑一遍」变成对 ci.yml 的契约) —— 新源码 scripts/ci_preflight.py 把工作流的 10 个步骤原样搬回本地跑,输出 reports/ci_preflight.json;v3_gate 长出第 33 面 A13,对外口径 32/32 → 33/33。为什么要它:186 连败里最贵的一课不是某个测试红了,而是本地全绿、云端全红 —— F 盘素材、客仓、交付包、CRLF 口径清单,本地都有,Linux runner 上一个都没有。本批不承诺「本地跑等于云端跑」,它承诺的是两张步骤表不许悄悄漂移:ci.yml 加了步骤而预检器没跟上,A13 当场打红。零新增依赖、零网络、除报告外不写盘。DOC 0.9.70 → 0.9.71;清单 250 → 251 条。 |
| 上一批次 | V77 账面一致性批(闸给自己补上一面「漏登记即红」的自覆盖) —— claims_gate 第 8 面 registry_coverage:6 个活体窗口 × 4 类模式(版本范围 / 徽章用例数 / 主套用例数 / 复算条数),窗口取锚点行 ±1 行,只扫「活体」不扫历史(历史副本按 timeboxed 原样保留,不删不改)。首跑红 6 处:中文 L158 V0.1 → V38、英文 L1069 V0.1 to V59 都被写死在活体时间线里且从未登记 —— 这正是 V75「建了七面闸、却漏改自家 README」的病理;两处改成回读式后 PASS / mismatch 0(totals 17-3-5-6)。同时撤销三处过期 PL 挂账(PL-1~PL-6 已在 V68/V69/V71/V72/V73 全部落地),并修掉闸里一处假绿:0.9.69 与 0.9.70 曾被 _int_of 抓首段数字归零判「相等」,版本升号对闸完全隐形,现改逐字符比。零新增依赖、零网络、不写盘。 |
| 上一批次 | V76 双口径收口批(PL-3 语义桥从「覆盖主路」改为「双口径并存」) —— 一份新源码 dual_caliber.py + 一份只读 data/dual_caliber_ledger.json,零依赖、零网络、除台账生成外不写盘。问题不是「桥接分没用」,而是一旦覆盖主路,实测 15 样本里 12 个归属改变(80%),V72 已推送的 18 仓归属账会被整体推翻。本批不改主路一个字节,改的是「第二口径怎么被读」。①主体:absorb_dual(spec) 主路可调用的双口径入口——base 逐字段等于 forge_absorb 现跑(重算、不缓存、不替换),bridge 逐字段等于 bridged_absorb 现跑;三态判定 bridge_adopted(换仓且有命中词)/ base_kept(同仓,分数可升)/ vacuum_kept(仍 0.0,登记在册)。②台账:15 样本逐条 base / bridge / 分值差 / 归因词 / 判定,totals zero_before 13 · zero_after 1 · changed 12 · adopted 12 · base_kept 2 · vacuum 1,每个数字与 V69 实测互证;self_digest 38adbb82332911b7(10,173 B,含末尾换行)。③七面可失败判据 + 十只金丝雀:base 漂移、bridge 漂移、判定漂移、采纳无词证、未登记真空(把有词证的样本谎报成零分)、自摘要、totals、行序错位、多增行、边界篡改——每面单独打红。④真空纪律:SG-MASTER-SONGJIANG 桥后仍 0 分,登记在册,不伪装已修复;认知主权 从 0.0000 升到 0.1096 但同落 C01,判 base_kept 而非 adopted —— 同仓不等于采纳。⑤四人组本轮实跑留档:check_emergence(0.85, 0.88, 0.80, 0.35) → emergence=true / order=0.569;约束审计 verdict=review,唯一激活约束是软约束「全量重评下游」。诚实边界:四人组「法」步的 convexity / governable 实测为浮点伪影(价值函数线性 ⇒ 二阶差分恒 0;5 组扰动里 1 组 +0.0 判 convex、4 组 −0.0 判 concave),不得用来背书可治理性;P2×P3 条件读常量 QUARTET_SIMS(0.7 ≥ 0.7 恒真),实际只有 P1 / P4 / order 三条件在判。18 仓远端重推属外部动作,本批停在 LOCAL_READY,只出确认表不 push。新增 test_v76_dual_caliber.py:42 例;主套 1835、redline 156、DOC 升至 0.9.69;仓内 .py 递归 156。⑥claims_gate._receipt() 把回执文件名写死成 reports/v75_receipt.json,于是本批 README 改对了仍读到 V75 的数——这正是本模块要治的病,现在还给了自己;改为按 reports/v<N>_receipt.json 里 N 最大者自动发现,verify 随之 PASS / mismatch 0,V75 的 50 例仍全绿。 |
| 上一批次 | V75 双注册簿闸门批(叙述位与外部规定版本升级为「可失败的只读注册簿」) —— 一份新源码 claims_gate.py + 一份只读 data/claims_registry.json,结构上只读、零依赖。问题不是「README 写错数字」,而是一个会变的事实被复制成静态文本副本,副本没有单一真源、没有核实时间、没有过期判据。本批不改任何数字,改的是「数字从哪来」。①doc_claims 15 条三段:current 9 条(徽章主套数 / 验收行 / 门禁行 / 红线文件数 / 清单条数 / DOC 版本,expected_from 一律指向批次回执或可计算值)、historical 3 条(V46-D tests-526 注记与 V73 快照,必带 as_of_batch + timeboxed,保留不删不判红)、narrative 3 条(环数 / 批次数 / 文件计数命令,一律回读式,写死即红)。②regulation_pins 6 条外部规定针(MCP 双版本 / 公卫规范第三版 2017 不含慢阻肺 / 公卫 2025 三文号 / 医保目录 2025 / 2025 版药典 / 等保 2.0),每条四必填(文号 / 版本 / 核实日 / 状态)+ 指针字段,只存文号指针不复制法规全文,status=unknown 是合法状态(对齐 V69 KNOWN_VACUUM)。③七面判据 + 七只金丝雀 7/7 全红:注册缺项、回读篡改、去时点盒、去核实日、核实日改旧、改 totals、删链位常量——每面都能单独打红且互不串扰。④README 双份按块 E 九处逐点到回读式:十一环改「环数以 claims_gate report 回读为准」并补 ring12–ring24 功能表(不写「二十四环」)、ls 顶层口径改 find 递归口径、V38/V60/77 条改台账回读、仓外 readme_claims_check.py 改为仓内 python -m holoaeon.claims_gate verify、V46-D 注记登记为 historical、MCP 补双版本、公卫标第三版、语义证据补三通道与虚高声明、仓集群补权威源口径(域仓 Gitee 唯一 remote,GitHub 为 V55 建仓快照,五地同头只适用 N19 主仓,域仓如实标「两地同头」)。⑤链位覆盖实测比投喂包多 1 —— 除包内 6 个 ring 模块外,holo_forge.py(V28 链式执行层第三环 ring3)亦缺常量,7 处一并补齐;check_chain_position 实测 29 个常量 / 24 个显式声明校验通过(多值列举与「(ring11 召回层)」类引用只要求常量在位,不判值)。⑥CI 第 10 道 Claims gate (A1 narrative + regulation pins),只加这一道。新增 test_v75_claims_gate.py:50 例;主套 1793、redline 155、DOC 升至 0.9.68;仓内 .py 递归 155。诚实边界:verified_at 是人写快照,本模块不联网核实规定现行有效,只判超期;面 2 保证「文档视图 == 批次回执记录」,不是「重跑一遍测试」;holo_forge 的 ring3 属历史链式执行层编号,与领域编排环并存,已在 lineage_note 写明非编号冲突 |
| 上一批次 | V73 史前大部头导学产品批(指针卡片 · 四部头 · 六面可失败) —— 一份新源码 prehistory_tutor.py,结构上只读。PL-6 把史前书库收成一套指针卡片:data/tutor_cards.json 共 42 张,分部头为 导学大典 14 / 计算机语言通史 5 / 周易 10 / 道藏 13;每卡只存 相对指针 + 行号 + 段落锚 + 200 字内短摘要 + 源文件 sha16。不夹带全文这句话是可判的:摘要超过 200 字即红——于是「没复制」成了能复算的事实,而不是一句承诺。零诊断零处方单独成一面:植入一个禁词,只让这一面变红,不允许被其它面的绿灯平均掉。四部头覆盖有下限(每部头 ≥3 张),所以「一本书抽得好」糊不过去。导学路径是单链,前后互指对称 + 成环即红。确定性是同一 digest 两跑一致,外加 self_digest 抓住建完再改。先证可失败再谈可信:金丝雀喂 6 份坏库——指针断、摘要 240 字、植入处方词、删掉一个部头、路径成环、报告被改——6/6 全红。真实站点 digest 9850bb6644a0954c,42 卡 / 4 部头;素材侧 5 件 / 440 文件 / 45.57 MB,只读、零搬运。新增 test_v73_prehistory_tutor.py:40 例;主套 1743、redline 154、DOC 升至 0.9.67 -> |
| 上一批次 | V72 18 仓判据注入包批(十八道闸、一份契约、两端各算一遍) —— 一份新源码 prehistory_injection.py,结构上只读。PL-5 给 18 个本地仓每仓塞一份三件包(ledger.json / prehistory_check.py / tests/test_cXX_prehistory.py),并把 CangGate.self_check 扩出一个 prehistory 快照——只读本包账、核对账路径与判据数、指针数,不 import 检查器(避循环)。每仓断言同一组六条:九件身份字段非空、仓号 ^C\d{2}$ 且与仓名前缀相符、github.com/…/holoaeon 锚点已声明、ring 恰为 19、密令一字不改、11 条机制指针齐全且无多余键;账在仓内跑一遍。先证可失败,再谈可信:金丝雀喂 6 种坏账——身份字段清空、仓名跨号错配、ring 改 18、密令多加一个句号、指针删一条、越界 import holoaeon——6/6 全红,所以这里「通过」是有意义的。注入后实测:18/18 仓 × 14 例 = 252,每仓判据 6/6、指针 11/11,零第三方依赖。主仓不复用这套实现:prehistory_injection.py 汇总 18 份账并独立再算一遍同一契约(两端验证,不是复制),失败面是聚合层真看得见的那些——仓数、仓号重复或断号、身份与仓名不符、schema 漂移、判据或指针缺额、绝对路径漏进台账、建完再改被自摘要抓出。零实现复制靠扫 import 语句与 sys.path 注入判定,不靠读全文——文本扫描会撞上自己的 pattern 字面量(检查器第一版就是这么把自己判红的),所以仓内用 AST、主仓侧按语句形状判。真实站点 digest 9265369717e254d5,18 仓 / 108 判据 / 198 指针。新增 test_v72_prehistory_injection.py:38 例;主套 1703、redline 153、DOC 升至 0.9.66 -> V71 史前矩阵族谱批(族谱写得出形状,写不出血缘) —— 一份新源码 prehistory_family.py,结构上只读、零搬运。实测推翻了策划案的前提:在严格骨架签名(相对路径 + 字节数排序后取 sha256[:16])下,216 个对象两两不同,域内同骨架族为 0。唯一的精确族是 14 对跨域双态——同一资产的「目录版」与「runtime zip 版」清单与大小完全一致,可判「双态」、不可判父子,故称 twin。于是族谱分两层:L1 精确 14 族 / 188 个单点;L2 近亲族 = 同域内文件名集合 Jaccard ≥ 0.50 的连通分量,实测 4 族,这才是母版/派生真正发生的地方(同门不同版:世殊医道 V4 对师叔值班分身 V3 = 0.714、旴江医派 20261018 对 20261020 = 0.578)。4 个近亲族里有 2 个正好是反证:冀派中医 对 吉派中医 0.933、燕京医派 对 颜德馨 0.950——不同对象、同一套模板,名字几乎一样。所以每个近亲族都带 evidence = name_jaccard>=0.50 与 content_verified = False,回执明写「未做任何字节级核验」。母版规则确定且写明:文件多 → 字节多 → 名字短 → 字典序,叫「同族骨架最完整者」,不叫「最早的版本」——历史先后需要时间戳,骨架给不出来。九道失败面可判 FAIL:目录/zip/文件/字节四项总数漂移、精确族数或单点数漂移、虚构血缘(族内签名不一致)、母版缺失、母版违约、孤儿差异(派生没有可复算差异)、差异复算不一致、近亲证据缺失或 content_verified 被翻成 true、绝对路径漏进签名、自摘要。新增 test_v71_prehistory_family.py:41 例;主套 1665,redline 152,DOC 升至 0.9.65 -> V70 挂账解题批(给每条挂账装一个「可以合法停在哪」的停止态) —— 一份新源码 prehistory_criteria.py,结构上只读、没有任何搬运函数。三件扯不清的账,这次都靠实测结账,不靠争论。①「194 还是 199」根本不是「+5」,而是 +7 / −2 的双向差:V1 严格判据选 194、V68 最简可复现判据选 199、交集 192、并集 201;并集这次带上了自己的指纹 set_sha16 = aae1c8aea9199ee7,口径定为 sha256("\n".join(sorted(names)))[:16]——这个口径是从数值本身反推出来的(11 种序列化穷举、2 种命中),所以「201」不再是一个只能信的数目,而是一个能自己算出来的数目。核心 194 与全集 201 并列登记;199 这一档从此只作对照留档,不再单独作唯一口径。②rc=0 不等于「全绿」:22 个 selftest 拆成三层证据——L1 静态 22 / L2 实跑 22 / L3 有断言且实跑 21(assert 合计 1410、min 0、max 96),唯一零断言实例(北京瑶医 198 行)登记为已知例外,此后任何新增零断言实例直接判 FAIL:空转的自检不再被平均数掩盖。③B4 剩的 9 项 / 119,157 文件 / 10,166.9 MB 全部停在 SUSPENDED:模块只出九行三态点单表(AUTHORIZED / EXCLUDED / SUSPENDED),没有搬运函数,于是「还没点单」是代码能表达的状态,而不是要靠人记住的一句话;S4(18 仓本地产物)保持 not_started。失败面:判据漂移(计数离开 194/199/201)、集合漂移(sha16 一动即红)、差集漂移(+7/−2 名单变化)、层级越权(L3 > L2)、未授权触碰(挂起项溜进第一批),外加自摘要(建完再改即抓)。新增 test_v70_prehistory_criteria.py:51 例;主套 1624、redline 151、DOC 升至 0.9.64 -> V69 领域语义桥批(把「零分兜底」换成一条敢承认自己不懂的桥) —— 一份新源码 prehistory_semantic_bridge.py,结构上只读。实测比策划案记的更糟:只喂项目名时,15 个史前资产里有 13 个对 18 仓全部打 0.0000,于是按仓名平票兜底到 C01 本源——这是兜底,不是归类。(策划案写「4 个零分」是另一个输入口径:它的样本带着描述与标签。两个数都真,但不能互比;回执把这条差异写成 feed_corrections,而不是悄悄挑好看的那个用。)桥没有改一行打分代码:ngram_dice 照旧、平票取小编号仓的规则照旧;它只在两侧各加一段显式、人写的词表文本——资产侧拿到自己名字里原本就有的区分性词(沙盘/法务/医派/数学),仓侧拿到该领域自有的同义词——于是同一个字符二元组 Dice 终于有了可比的共同词汇面。同一 15 样本上:零分 13→1、仓归属变动 12 个,剩的那 1 个是 SG-MASTER-SONGJIANG——一个不含任何领域词的纯英文代号,登记进 KNOWN_VACUUM 册,而不是化装成「已修复」。数学大师…数学史早课 不再落 C09 关系(改落 C05/C14,两者都更近)。四道失败面可判 FAIL:基线漂移(若实测 13 不再等于 13,说明输入口径被悄悄改了)、桥后零分超上限、残留零分不在真空册上、卡覆盖不够样本数;外加自摘要(建完再改即抓)与两跑一致性。词表是先验,不是理解——这句话同时写进模块 docstring 与回执,因为一条默默自称「看懂了」的桥,和 V67 删掉的 1e12 死常量是同一种病。新增 test_v69_semantic_bridge.py 43 例;主套 1573,redline 150,DOC 升至 0.9.63 -> V68 史前记忆打捞批(指针总账 · 骨架契约 · 把「断言」数清楚)—— 两份源码,都以"结构上只读"为前提。①prehistory_ledger.py 把 39.7 GB 的史前面摊成指针账:590 个顶层条目原样继承 V1 分类清单,只做一处授权改判——七项活体同步/镜像设施(vector-sea,即五地同步仍在读的 S4 权威源,加云端镜像、阿赖耶识镜像、同步状态、本地库及其快照)整体移出史前列:把活设施登记成"史前遗产"是账目污染,比漏登更坏。第一批作业面仍是 B1+B2+B3+B5 = 304 项 / 10,174 文件 / 1,944 MB,逐项算 meta_sha16(对"相对路径|字节数"排序后取哈希);目录指纹故意叫 metadata 指纹——它只能证明"清单没变",永远不能证明"字节没变",两种指纹在字段名上就分开。校验留四个独立失败面(磁盘缺件 / 指纹漂移 / 活设施混进 B 类 / 仓库路径漏进史前账),外加自摘要,建账后再改任何一处都会被自己的 sha256 抓出来。②prehistory_contract.py 把"22 个 runtime 长得像"从形容词落成断言,只用 ast 解析、不执行任何史前代码:22 实例 / 591 entries / 481 py,19 位同构文件名 22/22 全覆盖,共同函数签名 65 个分布在 16 个文件。策划案有两条经不起实测:签名是 65 不是 21(21 只是五个文件的抽样口径),"每实例 101 个 assert" 不成立——min 0 / max 96 / 合计 1410,而 北京瑶医医院项目管理…_runtime_20261018.zip 的 selftest.py 有 198 行、G1–G11 标记齐全、assert 为 0:有自检之形、无自检之实。这一例登记为已知例外,此后任何新增的零断言实例直接判 FAIL——契约把空转摆出来,而不是平均掉。新增 test_v68_prehistory_ledger.py:54 例;主套 1530、redline 149、DOC 提升 0.9.62 -> V67 契约对账与仲裁批(白名单对账 · 优先级裁决 · 把"验过"的边界说清楚)—— 三份源码。①resonance_supplement.py 删掉死常量 NO_INFO_P = 1.0e12:V66 起改用信息量表示 Y=1/P,Y=0 本身就是真·无信息先验,那个"假装很大"的有限值反而磨掉 2e-6 有效位数,且全仓已无一处引用;原位留一段 # V67 清理 说明,因为注释比陷阱便宜。②topology_resonance_pack.py 补上两种"反向病"的显式仲裁——混乱侧(熵>0.95 或 秩序<0.3)与僵死侧,用 REORGANIZATION_PRIORITY = "stagnation_first" 定口径(可一键改 disorder_first / all):耗散结构的前提是"有流",零流量下自组织不可能发生,而混乱侧至少还有噪声可被结构捕获。实质修正是:单点快照分不开"高效稳定"与"僵死"——全成功时 σ=2.0、熵=0,真僵死系统长得一样;故僵死改为时间窗证据**,新增可选入参 streak,只有 streak ≥ 5 且 entropy ≤ 0.10 才成立;不传 streak 则 stagnation_side = None(未评估,不假装评估过),σ < sigma_min 降级为weak_flow_signal 弱信号而不再单独判僵死。旧键 needs_reorganization 语义不动(只表混乱侧),另给 needs_reorganization_any 供下游一次判。③emergence_chain 现在**消费** consistency_gate 与 entropy_production——经 dissipation_view / consistency_view / advisories 三个新键读取并展示;但布尔判定仍只有 quartet 一个真源(P2×P3≥0.7 / P1≥0.6 / P4≤0.4 / order>0.30):若把一致性门与熵产率也接进判定,等于给"涌现"开第二、第三真源,口径一分歧就互斥。④dify-workflow 契约由"读到的"向"跑通的"推进:本机回环(自己起的 http.server,不是** Dify 实例——本机五次探端口全 False、无 docker、无 app token)真发 HTTP,用仓内 build_request 造的真实请求体验 15/15——路径 /v1/workflows/run、Bearer 头、恰三键 body、缺省 blocking、响应 11 字段、缺 user 返 400 invalid_param、streaming 开关;故 verification 变为 source_code_read+local_contract_loopback,并配一段"这不证明什么"的明文边界(Dify 官方可用性、真实工作流输出、真实配额与错误码触发)。新增 test_v67_contract_and_arbitration.py:把 8 个函数的返回键锁成精确集合(用户点的是 7 个,实测被消费的是 8 个——rff_features 是 V66 新增的单点版、被 hilbert_project 直接依赖)、锁 13 键/10 键的卡尔曼行级契约、锁仲裁四分支与口径切换、锁 Dify 静态契约:43 例;主套 1476、redline 147、DOC 提升 0.9.61 -> **V66 判据回写原包批(7 把尺子搬进 topology_resonance_pack.py 本体,软内核整段换掉——原包改为从 resonance_supplement.py 取数学,单一真源、不留第二份会漂移的副本;随后逐位换芯:①mandala_fanout 删掉 _det_hash % 10 > 2 那把骰子,改成八方位 Hadamard 正交变换(往返误差 0、正交偏差 0、Parseval 比 1.0、直流等于均值 4.5;八方位齐步时只剩直流项,有效扇出降为 0——没有分化就没有扇出,判据可失败);②kalman_state 把写死的单源标量滤波改成「预测步 + 标准 NIS」(NIS =(z−x⁻)²/(P⁻+R),χ²(1) 95% 门限 3.841)+ 新息门控:故意塞入 z=30 的离群源,NIS 557.99 判拒,融合值保持 10.371429,不再被拉到 15.786207;把门控关掉则精确复现旧口径 15.786207——证明旧差距来自缺预测步与缺门控,而不是数据不同;kalman_fuse 同时改为信息量表示 Y=1/P,q=0 时序贯结果与批量逆方差加权精确相等(1e-9 容差),check() 仍 7/7 PASS;③bayesian_guard 把写死的 0.5 阈值换成贝叶斯风险最优的代价敏感阈值 C_fp/(C_fp+C_fn)=1/11=0.090909(network 后验 0.6136 仍 BLOCK、read 0.0145 ALLOW),并新增证据路径走 Wald 序贯检验(A=2.890372/B=-2.251292,第 5 步停);④hilbert_project 把假投影(φ_k=1/√(k+1)、v[k%n]、范数随维数发散)换成随机傅里叶特征,E[‖z‖²]=1,投影范数不再随 N 膨胀;⑤dissipation_measure 补上熵产率 σ=ΣJ_iX_i(力取 [p_ok, p_fail],走 Onsager 对称 + Sylvester 正定,实测 σ=1.416、主子式 [2.0, 1.91])与耗散充分性判据,Shannon 熵与秩序度原键不动;⑥resonance_node 补上真序参量 r= |
| 上一批次 | V68 史前记忆打捞批(指针总账 + 骨架契约) —— prehistory_ledger.py 原样继承 V1 清单(590 顶层,仅一处授权改判:七项活体同步/镜像设施移出史前列);prehistory_contract.py 只用 ast 读 22 个 runtime(16 文件 65 个共同签名、assert 0–96 合计 1410、登记 1 个零断言实例);54 例,主套 1530,redline 149,DOC 0.9.62;再上一批 —— **V65 拓扑共振与耗散验证域·外部补齐批(7 位「只有承载者、内核偏软」的机制换上真判据——resonance_supplement.py 落 9 条原理锚(Prigogine 1977 诺贝尔演讲讲耗散结构、皇家学会 50 周年综述讲最小熵产定理、Bak-Tang-Wiesenfeld 1987 自组织临界、Kalman 1960、NIS 一致性检验、Wald 1945 序贯概率比、Rahimi-Recht 2007 随机特征、Kuramoto 1975)+ 12 条上游台账(Kalman 讲义 19480★/filterpy 3899★/TopologyLayer 431★/torchhd 394★/ripser.py 347★/沙堆 7★/Kuramoto 9★/随机共振 19★),一律 method_only 或 readonly_reference,一行代码未抄;再上 7 个可实测判据:①C1/R8 熵产率 σ=ΣJ_iX_i 走 Onsager 对称 + Sylvester 正定(实测 σ=2.55、主子式 [2.0, 1.91]、σ(tX)=t²σ(X) 在 t=0.5/2/3 比值均为 1.0);②R2 曼陀罗一拖八从「哈希假装成功」改成八方位上的 Hadamard 正交变换——往返误差 0、H8·H8=8I、中心等于均值(载荷 1..8 时 4.5)、八方位齐步时只剩直流项、Parseval 比 1.0;③R5 卡尔曼改信息式 P^-1=ΣR_i^-1 吃异构 R(融合 10.371429、P 0.190476、权重 0.190/0.762/0.048、NIS 0.116/0.038/0.078 全在 χ² 3.841 内)+ 序贯等于批量 + 故意塞离群源能被判出;④R6 守卫改 Wald 序贯检验,真门限 A=2.890372/B=-2.251292,H1 在第 5 步停、H0 在第 9 步停,不再等固定窗口;⑤R7 希尔伯特投影换随机傅里叶特征(Bochner),映射变成真的核近似:D=16 时 MAE 0.198 降到 D=256 时 0.035,固定种子可复现;⑥X2 共振有了序参量 r= |
| 上一批次 | V86 内核基因 DNA(三条内核基因落成可失败断言,经络不再写死 10 碎片) —— ①holoaeon/genome_dna.py 把三条内核基因做成可失败判据:G-DNA-01 生长≠分裂(Δedges>0 ∧ Δcopies=0 ∧ 协议族同源 ∧ 封存同源 → 生长;Δcopies>0 → 分裂)、G-DNA-02 分化≠分裂(表型专化 ∧ 三同源 DIFF=0:25 机制轴 / 心跳 70·768·17 / SEAL;θ 禁 1.0、未标定前 theta=null、禁宣称 100% 同源)、G-DNA-03 鲁棒性真义(R1 局部失效可计量、R2 扩展挤压下 spec 字节相同 + genome_hash DIFF=0 + 格数守恒、R3 三同源守恒);②meridian_link.py 削平六处旧皮——len(rows)!=10 上限守卫改为 BASE_FRAGMENT_COUNT=10 下界并与 genome_dna.registered_fragments() 合并、建表动态化、check_linkage 用 expected = 碎版数 × 25、meridian_summary / build_ledger 的 10 改为 len(load_fragments());③新增 scripts/asset_census.py 实测 顶层 616 项(418 目录/198 文件)→ 净候选 201 / V3 族 149 / 顶层 zip 69 / 四件齐备 0(投喂包"四件齐备"的前设在本机不成立——如实登记,不悄悄改口径);④新增 scripts/plugin_contract.py 按一份八面契约装配 3 个试点,实测 θ = 0.592(取双态真试点相似度最小值;第 3 试点实测 0.368,判「分裂/孤儿插件」,dual_state=false,如实不装配),双态真 2/3;⑤注册簿 append 3 条碎片,经络实盘 300 格(12 × 25),三项 PASS;⑥A14 恒真金哨由"常量对常量"改为读 DNA 母本 seal 对账(v3_gate 33/34 → 34/34),相对导入修复后红线 159 文件 0 违规;⑦DOC_VERSION 不顺延——它在 spec.py 内,字节相同即第一判据;⑧V85 治理批未执行(无技术依赖、用户本轮要求先同步),如实登记在 honest_boundaries,不化妆。主套 2243 → 2285 · 新增 test_genome_dna 42 例 · V62/V71 测试去写死(注入空 extra_fragments 基线 + pf.EXPECTED_* 单源)· 清单 341 条不变(新增件按 V41-A 惯例只以 git 跟踪留痕) |
English historical ledger (moved from README.md)
| Batch | Content |
|---|---|
| Previous batch | V84 always-true criteria standing golden watch (tautology_scan A-caliber scanner + allowlist ledger + v3_gate face 34 A14) — V82/V83 fixed the two known always-true criteria by hand; nothing stopped the next one. This batch turns the scan into a standing gate: ①new holoaeon/tautology_scan.py (pure AST, zero execution, zero new dependency, py3.8-compatible) detects the four constant-vs-constant shapes (tautology_eq / dead_always_true / dead_always_false / membership_in_const); ②data/tautology_allowlist.json registers the 8 legitimate uses (invariant self-check / config-constant dispatch / dict presence check), each carrying a liveness source — the allowlist is a registry, not a permission list (missing liveness / stale entry / digest mismatch all fail); ③scripts/v3_gate.py grows face 34 A14 with three falsifiable criteria unregistered_tautology · allowlist_drift · scanner_selfcheck; ④the scanner self-check absorbs both parties' v1 lessons (25 chain-compare false positives on the quartet side, 16 constant-environment false positives on the execution side); ⑤8 canaries C1–C8. The static caliber only judges B-caliber (literal constant comparison): an unregistered hit is a pending review, not a proven defect; A-caliber (does the criterion actually constrain its input?) still needs a human read. ci.yml byte-identical (6th consecutive batch) · DOC 0.9.77 → 0.9.78 |
| Latest batch | V83 same-shape always-true disposal #2 (real p23 input for topology_resonance.quartet_tensor) — V82 fixed the always-true criterion in quartet_director and named this second same-shape case in honest_boundaries; this batch fixes it. (1) quartet_tensor gains a real p23 input; the default reads the function's own _cosine(vecs["P2"], vecs["P3"]) (vecs = [s, 1-s], the same infrastructure that already produces res_pairs) — a genuine measurement, not a convention, and different from, yet coexisting with, V82's `1 - |
| Latest batch | V82 P1-B failed-criteria disposal (two always-true / artifact criteria move from disclosure to disposal) — V81 only disclosed; this batch actually fixes two source-level dead criteria. (1) advisor_gradient_flow's governable leaves its boolean squeeze and becomes True / False / None, with a relative curvature tolerance `tol = 1e-9 * max(1.0, |
| Previous batch | V80 cloud-vs-local environment delta made falsifiable (ENV_DELTA registry) — "all green locally, 46 fewer tests executed in the cloud, same green" stops being a matter of luck: claims_gate grows face 11 env_delta_consistency plus data/env_delta_registry.json (D01-D11), with three falsifiable criteria unregistered_guard (every environment guard must be attributed to an entry), parity_disclosed (the ran / skipped / executed deltas are explained entry by entry per suite, and both READMEs disclose local and cloud numbers side by side), and digest_drift (each line anchor must sit within +-10 lines of its probe token; one byte edited without re-sealing the self digest is red), with eight canaries each asserting a red. Proved red first (unregistered state: 25/25 guards red, mismatch 29), then turned green. Cloud py3.12 (run 43540948562 after the V80 push) Ran 2031 / skipped 33 and contract 184 / 11 versus local 2036 / 3 and 184 / 0 are now fully named (F-drive assets 30 skips + 5 undiscovered, guest repo 10, delivery package 1, out-of-repo module 3 both sides); already-fixed items (CRLF, timezone) are lineage only and may not be borrowed to explain a live delta. No new module, ci.yml byte-identical (the face grows inside the existing step); face numbering is read back from the report, never hardcoded |
| Previous batch | V78b CI gates formally sealed (the recompute gate and the v3 domain gate enter ci.yml, 10 → 12 steps) — P0-A line-ending normalization: check_sums hashes text files on an LF byte caliber (binaries raw per the .gitattributes binary lines, the single rule source, never hardcoded in the gate), so a Windows CRLF working tree and a Linux checkout verify the same manifest — "local green = cloud green" stops being a slogan and becomes a judgment. P0-B manifest hygiene: five runtime artifacts removed (three untracked reports/daily_* + two gitignored feishu fetch dumps — none of them exist on a cloud checkout) and the gate grows untracked / ignored machine checks, so a runtime artifact sneaking into the manifest is machine-blocked instead of caught by hand-auditing; the rebuild tool build_b8_dual_sync.release_files likewise collects only git-tracked files. P0-C: the Manifest integrity (recompute) and v3 gate (33 faces) steps formally enter ci.yml; the ci_preflight contract grows 10 → 12 faces in lockstep (the A13 two-way alignment check scales with it, and test_v78_ci_preflight asserts the new counts). claims_gate._receipt now picks the latest receipt by the receipt's own doc_version (the letter-suffixed batch V78b is chronologically after V81; sorting by the N in the filename would point at the wrong receipt). SHA256SUMS rebuilt: 251 → 332 entries (five removed + 86 tracked files that V-series batches had added but the 11-02 manifest never caught up with). DOC 0.9.72 → 0.9.73 (four coupled sites in the same batch); local all-green: main suite 1931 OK · v3_gate 33/33 · claims nine faces · recompute 332 entries / 0 mismatches. |
| Previous batch | V81 regulation-contract semantics + a general freshness contract (claims_gate grows a face that catches "registered, but not what the runtime does") — the debt was not "README missed a version": README.md said MCP was single-version 2024-11-05, mcp_adapter.py's PROTOCOL_VERSION default was the same, yet the registry recorded 2026-07-28 as current — three sources disagreeing while every gate stayed green. No new module; ci.yml byte-identical (sha256[:16] b743704ca8c307f7 before and after). The new face regulation_semantic_consistency (ordinal read back via python -m holoaeon.claims_gate report) has three criteria: ① every claim_site is re-read and an unregistered version token raises claim_site_drift; ② runtime_pointer + runtime_role/runtime_value register the constant that actually runs, and legacy must carry superseded_by; ③ one freshness_contract schema covers both external pins and internal measured baselines — risk_tier grades fast 90 / annual 180 / stable 365, any source_sha16 drift is red, and change_probe.mode only accepts manual_review_only (an offline project does not claim it verified externally). Registry: 6 pins extended with six fields + 4 internal baselines whose measured_at matches each file's mtime — nothing back-dated. Red first, then green: on the unextended registry the face measured 42 red / pins_with_contract=0; 8 canaries 8/8 red, with C1 and C2 landing on the real R1 (README.md#L364, single version) and the real R2 (runtime_pointer missing). After registration: PASS / mismatch 0. New test_v81_regulation_semantic.py 53 cases (ten groups); main suite 1874 → 1931; redline 156; DOC 0.9.71 → 0.9.72. Boundary: this batch only makes the decoupling falsifiable — the PROTOCOL_VERSION default is deliberately not moved to 2026-07-28 (compatibility is a carbon-level decision); verified_at stays a human snapshot, and the gate judges drift and age, not whether a regulation is still in force. |
English historical ledger (moved from README.md)
| Batch | Content |
|---|---|
| 上一批次 | V78a CI preflight gate (the local run becomes a contract against ci.yml) — new scripts/ci_preflight.py replays the workflow's ten steps locally and writes reports/ci_preflight.json; v3_gate grows face A13, moving the gate basis 32/32 → 33/33. Why it exists: the most expensive lesson of the 186-run failure streak was not a red test but local green with cloud red — F-drive material, guest repos, delivery packages and CRLF-caliber checksums all exist locally and none of them exist on a Linux runner. This batch does not promise that the local run equals the cloud run; it makes the drift between the two step tables impossible to pass unnoticed: a step added to ci.yml that the preflight does not claim turns A13 red on the spot. Zero new dependencies, zero network, no writes beyond the report. DOC 0.9.70 → 0.9.71; SHA256SUMS 250 → 251 entries. |
| 上一批次 | V77 ledger-consistency batch (the gate grows a face that catches its own omissions) — new face 8 registry_coverage on claims_gate: six live windows × four patterns (version ranges / badge counts / main-suite counts / recompute rows), each window anchored to one line ±1, scanning live text only — historical copies stay verbatim under timeboxed instead of being deleted. First run came up red in 6 places: Chinese L158 V0.1 → V38 and English L1069 V0.1 to V59 were hardcoded inside live timeline windows and had never been registered — exactly the pathology V75 left behind when it built a seven-face gate and forgot its own README. Both rewritten as readback → PASS / mismatch 0 (totals 17-3-5-6). Three stale PL backlogs retracted (PL-1..PL-6 all landed in V68/V69/V71/V72/V73), and one false green removed: 0.9.69 vs 0.9.70 used to be zeroed by _int_of and judged equal, so a version bump was invisible to the gate; now compared character by character. Zero new dependencies, zero network, no writes. |
| 上一批次 | V76 dual-caliber closeout (PL-3's bridge stops overwriting the mainline and becomes a second caliber read side by side) — one new source dual_caliber.py plus a read-only data/dual_caliber_ledger.json; zero dependencies, zero network, no file writes outside ledger generation. The defect is not that the bridge score is useless: the measurement says overwriting the mainline moves 12 of 15 attributions (80%), which would overturn the attribution account of the eighteen repos already pushed in V72. This batch changes not one byte of the mainline; it changes how the second caliber is read. (1) absorb_dual(spec) is the mainline-callable dual entry: base equals forge_absorb field by field on every call (recomputed, never cached, never substituted), bridge equals bridged_absorb field by field, and the decision is one of three states — bridge_adopted (repo moved and matched words exist), base_kept (same repo, score may rise), vacuum_kept (still 0.0, registered). (2) The ledger records 15 samples with base / bridge / score delta / attribution words / decision, totals zero_before 13 · zero_after 1 · changed 12 · adopted 12 · base_kept 2 · vacuum 1, each figure cross-checked against V69's measurement; self_digest 38adbb82332911b7 over 10,173 bytes (trailing newline included). (3) Seven FAIL-able faces plus ten canaries: base drift, bridge drift, decision drift, adoption without word evidence, an unregistered vacuum (a sample with matched words falsely reported as zero), the self digest, totals, row-order swap, an extra row, and a tampered boundary list — each face goes red alone. (4) Vacuum discipline: SG-MASTER-SONGJIANG still scores 0 after the bridge and stays registered, never dressed up as repaired; 认知主权 rises from 0.0000 to 0.1096 inside the same repo C01, so it is base_kept, not adopted — same repo is not the same as acceptance. (5) The quartet was actually executed this round: check_emergence(0.85, 0.88, 0.80, 0.35) gives emergence=true / order=0.569, and the constraint audit returns verdict=review with the only active constraint being the soft "re-evaluate every downstream consumer". Honest bound: the quartet's convexity / governable reading is a floating-point artifact (the value function is linear, so its second difference is identically zero; 5 perturbations gave 1 +0.0 convex and 4 −0.0 concave), so it must not be used as evidence of governability; the P2×P3 condition reads the constant QUARTET_SIMS (0.7 ≥ 0.7 always true), leaving only P1 / P4 / order actually judging. Re-pushing the eighteen repos is an external action: this batch stops at LOCAL_READY and only ships a confirmation table. New test_v76_dual_caliber.py: 42 cases; main suite 1835, redline 156, DOC 0.9.69, recursive package .py 156. (6) claims_gate._receipt() hard-coded reports/v75_receipt.json, so this batch still read V75's numbers after the README was fixed — the exact disease the module was built to cure, handed back to itself; it now auto-discovers the largest reports/v<N>_receipt.json, after which verify reads PASS / mismatch 0 and all 50 V75 cases stay green. |
| 上一批次 | V75 dual-registry claims gate (narrative and regulation copies become a falsifiable read-only registry) — one new source claims_gate.py plus a read-only data/claims_registry.json; structurally read-only, zero dependencies. The defect is not that README carried a stale number: a changing fact was copied into static prose with no single source, no verification date and no expiry criterion. This batch changes no number — it changes where numbers come from. (1) doc_claims 15: current 9 (badge suite / acceptance row / gate row / redline files / checklist entries / DOC version, all expected_from resolving to the batch receipt or a computed value), historical 3 (the V46-D tests-526 note and the V73 snapshot, which must carry as_of_batch + timeboxed and are kept, not deleted), narrative 3 (ring count / batch count / file-count command, readback-only, hard-coded means red). (2) regulation_pins 6 (MCP dual-version / PHC norm 3rd ed. 2017 excluding COPD / PHC 2025 circulars / NHSA catalogue 2025 / pharmacopoeia 2025 / MLPS 2.0) with four required fields plus a source pointer — pointer only, never the regulation text; status=unknown is a legal state. (3) Seven falsifiable faces with 7/7 canaries red, one per face, no cross-talk. (4) README x2 rewritten to readback form per block E; ring count never hard-coded (ring12–ring24 map added), ls top-level replaced by recursive find, V38/V60/77-count replaced by ledger readback, the outside-repo checker replaced by python -m holoaeon.claims_gate verify, V46-D note registered historical, MCP dual-version, PHC 3rd edition, semantic evidence with the inflated-P@1 caveat, and cluster authority declared (domain repos: Gitee sole remote, GitHub is the V55 snapshot; five-site same-head applies to N19 only). (5) holo_forge.py (V28 chain-execution ring3) also lacked CHAIN_POSITION; 7 modules were completed, giving 29 constants / 24 declared-and-verified. (6) CI gate #10 Claims gate. New test_v75_claims_gate.py: 50 cases; main suite 1793, redline 155, DOC 0.9.68, recursive package .py 155. Honest bound: verified_at is a hand-written snapshot — this module never checks currency over the network, only staleness; face 2 guarantees "document view == batch receipt", not "tests re-run". |
| 上一批次 | V73 prehistory-tutor (pointer cards, four great books, six falsifiable faces) — one new source file, read-only by construction: prehistory_tutor.py. PL-6 turns the prehistoric library into a pointer card set: data/tutor_cards.json holds 42 cards across four corpora — tutor canon 14 / computer-language history 5 / Zhouyi 10 / Daozang 13 — and each card stores only a relative pointer, a line number, a section anchor, a **<=200-character** summary and a sha16 of the source file. No full text is copied: a summary over 200 characters fails, which is what makes "no copying" measurable rather than promised. Zero diagnosis and zero prescription is its **own** criterion — a planted term goes red on that face alone and cannot be averaged away by other greens. Four-corpus coverage has a floor (>=3 cards each), so a library cannot pass by sampling one book well. Study paths are single chains with prev/next symmetry plus a cycle check. Determinism is two runs over one digest plus a self_digest that catches edits made after the build. Measured before trust: a canary fed six bad libraries (broken pointer, 240-character summary, implanted prescription word, one corpus deleted, a cycle, a tampered report) and 6/6 went red. Digest 9850bb6644a0954c over 42 cards / 4 corpora; the material side is 5 artefacts / 440 files / 45.57 MB, read only and never moved. New test_v73_prehistory_tutor.py: 40 cases; main suite 1743, redline 154, DOC raised to 0.9.67 -> |
| 上一批次 | V72 prehistory-injection (eighteen gates, one contract, two independent readings) - one new source file, read-only by construction: prehistory_injection.py. PL-5 puts a three-file injection pack into each of the eighteen local repos (ledger.json / prehistory_check.py / tests/test_cXX_prehistory.py) and extends CangGate.self_check with a prehistory snapshot that reads the pack's own ledger, checks the criterion and pointer counts, and deliberately does not import the checker (no cycle). Each pack asserts the same six criteria - nine identity fields non-empty, cang id matching ^C\d{2}$ and the repo prefix, a github.com/.../holoaeon anchor declared, ring exactly 19, the seal verbatim, and eleven mechanism pointers with no extras - and each repo runs them against its own ledger. Falsifiability was established before trust: a canary fed six bad ledgers (identity blanked, repo swapped across cangs, ring 18, seal with a trailing stop, one pointer deleted, an out-of-bounds import holoaeon) and 6/6 went red, so a green criterion here carries weight. Measured after injection: 18/18 repos x 14 cases = 252, every repo 6/6 criteria and 11/11 pointers, zero third-party dependencies. The main repo reuses none of that code: prehistory_injection.py aggregates the eighteen ledgers and recomputes the same contract independently - a two-ended verification, not a copy - and its failure faces are the ones an aggregate can actually see: repo count, duplicate or gapped cang ids, identity mismatch, schema drift, criteria or pointer shortfall, an absolute path leaking into the report, and a self digest that catches edits made after the build. Zero implementation copying is judged by scanning import statements and sys.path injection rather than by reading file text, because a text scan trips on its own pattern literals - the first version of the checker failed itself for exactly that reason, which is why the rule is AST-shaped inside the repos and statement-shaped here. Real-site digest 9265369717e254d5 over 18 repos / 108 criteria / 198 pointers. New test_v72_prehistory_injection.py: 38 cases; main suite 1703, redline 153, DOC raised to 0.9.66 -> V71 prehistory-family-tree (the tree writes down the shape, never the blood) — one new source file, read-only by construction: prehistory_family.py. Measurement overturns the planning premise: under an exact skeleton signature (relative path + byte size, sorted, sha256[:16]) all 216 objects are mutually distinct — zero same-skeleton families inside a scope. The only exact families are 14 cross-scope twins: one asset's directory form and its runtime-zip form share listing and sizes exactly, which licenses the word twin and not the word parent. The tree is therefore two-layered: L1 exact (14 twins / 188 singletons) and L2 near kin — a connected component at name-set Jaccard >= 0.50, measured 4 families, and the only place where master-vs-derivative actually happens (same lineage, different versions: 世殊医道 V4 vs duty-avatar V3 at 0.714, 旴江医派 20261018 vs 20261020 at 0.578). Two of those four are the counter-evidence that skeleton similarity is not descent: 冀派中医 vs 吉派中医 score 0.933 and 燕京医派 vs 颜德馨 0.950 — different assets cut from one template, names almost identical. So every near family ships evidence = name_jaccard>=0.50 and content_verified = False, and the receipt states plainly that no byte-level check was run. The master rule is deterministic and written down: more files -> more bytes -> shorter name -> lexicographic, and it is called "the most complete skeleton in the family", never "the earliest version" — history needs timestamps, which a skeleton cannot supply. Nine FAIL-able faces: directory / zip / file / byte totals drift, exact-family or singleton count drift, phantom lineage (a member's signature disagrees with its family), missing master, master-rule violation, orphan diff (a derivative with no reproducible difference), diff not reproducible on re-read, near evidence missing or content_verified flipped to true, an absolute path leaking into a signature, and a self digest. New test_v71_prehistory_family.py: 41 cases; main suite 1665, redline 152, DOC raised to 0.9.65 -> V70 discharge-plan execution (every backlog item gets a legal place to stop) — one new source file, read-only by construction: prehistory_criteria.py. Three arguments were settled by measurement instead of by opinion. (1) The 194-vs-199 question was never a +5 — it is a +7 / -2 two-way difference. The V1 strict regex selects 194, the V68 simplest reproducible one selects 199, the intersection is 192, and the union is 201, which now carries its own fingerprint set_sha16 = aae1c8aea9199ee7 defined as sha256("\n".join(sorted(names)))[:16] — the convention was reverse-engineered from the value itself (eleven serialisations probed, two agree), so "201" stops being a number you must trust and becomes one you can recompute. Core 194 and full-set 201 are registered side by side; the 199 tier survives only as a cross-reference and is never the sole authority again. (2) rc = 0 is not "all green" — the 22 selftests split into three layers: L1 static 22 / L2 executed 22 / L3 has-asserts-and-executed 21 (asserts sum 1410, min 0, max 96), and the one zero-assert instance (Beijing Yaoyi, 198 lines) is a registered exception that fails the gate if a new zero-assert sample ever appears, so an empty self-test can no longer be averaged away. (3) B4's remaining 9 items / 119,157 files / 10,166.9 MB stay SUSPENDED: the module ships a nine-row three-state sheet (AUTHORIZED / EXCLUDED / SUSPENDED) and contains no move function at all, which makes "not yet decided" a state the code can express rather than a note somebody has to remember; S4 (the 18-repo local products) stays not_started. Failure faces: criteria drift (a count leaves 194/199/201), set drift (the sha16 moves), diff drift (the +7/-2 rosters change), layer over-claim (L3 > L2), unauthorized touch (a suspended item slips into the first batch), plus a self digest that catches post-hoc edits. New test_v70_prehistory_criteria.py: 51 cases; main suite 1624, redline 151, DOC raised to 0.9.64 -> V69 semantic-bridge (a bridge that states out loud what it cannot know) — one new source file, read-only by construction: prehistory_semantic_bridge.py. The measured problem is worse than the planning memo recorded: with name-only input, 13 of 15 prehistoric assets score exactly 0.0000 against all 18 cang identities and are therefore tie-broken onto C01 purely by repo name — a fallback, not a classification. (The memo's "4 of 15" is a different input discipline: its samples carried description and tags. Both numbers are real, they are not comparable, and the receipt records the correction rather than quietly picking the flattering one.) The bridge changes no scoring code: ngram_dice still does the work and the tie-break is still the smallest repo name; it only appends text on both sides from an explicit, human-written lexicon — the asset side receives the discriminating words already present in its own name (沙盘 / 法务 / 医派 / 数学), the cang side receives the synonyms that domain owns — so the same character-bigram Dice finally has a shared vocabulary to compare. On the same 15 samples: zero-score falls 13 -> 1, cang assignment moves on 12, and the one remaining zero is SG-MASTER-SONGJIANG — a bare English code name carrying no domain word at all, registered as KNOWN_VACUUM instead of being dressed up as repaired. 数学大师…数学史早课 no longer lands on C09 关系 (it now goes to C05 / C14, both defensibly closer). Four failure faces are FAIL-able: baseline drift (if the measured 13 stops being 13, the input discipline silently changed), zero-after above cap, any residual zero missing from the vacuum register, card coverage short of the sample set — plus a self digest that catches post-hoc edits and a two-run determinism check. The lexicon is a prior, not comprehension; that sentence sits in the module docstring and in the receipt, because a bridge that silently claimed to understand would be the same disease as the 1e12 constant V67 deleted. New test_v69_semantic_bridge.py: 43 cases; main suite 1573, redline 150, DOC raised to 0.9.63 -> V68 prehistory-ledger (the pointer ledger, the skeleton contract, and the honest count of asserts) — two source files, both read-only by construction. (1) prehistory_ledger.py turns the 39.7 GB prehistoric surface into a pointer ledger: 590 top-level items classified by inheriting the V1 list verbatim, with exactly one authorised reclassification — the seven live sync/mirror facilities (vector-sea, which is the authoritative S4 source the five-site sync still reads, plus the cloud mirror, the alaya mirror, the sync state, the local library and its snapshots) leave the prehistoric column, because filing a live facility as a dead artefact pollutes the account rather than merely omitting it. The first working surface stays B1+B2+B3+B5 = 304 items / 10,174 files / 1,944 MB with per-item meta_sha16 over sorted (relative path, byte size); the directory fingerprint is labelled a metadata fingerprint on purpose — it proves the listing has not changed and never that the bytes have. Verification exposes four independent failure faces (missing on disk, fingerprint drift, live infra filed as prehistory, repo paths leaking in) plus a self digest that catches any post-build edit. (2) prehistory_contract.py states what "the 22 runtimes look alike" actually means, from ast parsing only — no prehistoric code is executed: 22 instances / 591 entries / 481 py, the 19 isomorphic file names present in 22 of 22, and 65 shared function signatures across 16 files. Two claims in the planning memo do not survive measurement: shared signatures are 65, not 21 (21 is the sampling figure of five files), and "101 asserts per instance" is false — min 0, max 96, sum 1410, and 北京瑶医医院项目管理…_runtime_20261018.zip ships a 198-line selftest with all eleven G-groups and zero asserts, the shape of a self-test carrying none of its substance. That sample is registered as a known exception, and any new zero-assert instance now fails the gate — the contract exposes the emptiness instead of averaging it away. New test_v68_prehistory_ledger.py: 54 cases; main suite 1530, redline 149, DOC raised to 0.9.62 -> V67 contract-and-arbitration (the white-list audit, the priority ruling and the honest edge of "verified") — three source files. (1) resonance_supplement.py drops the dead NO_INFO_P = 1.0e12: nothing referenced it since V66 replaced the pretend-huge-number trick with the information form Y = 1/P, where Y = 0 IS the true no-information prior, and the finite 1e12 cost 2e-6 of precision; a # V67 cleanup note stays in its place, because a comment is cheaper than a trap. (2) topology_resonance_pack.py gains an explicit arbitration between two opposite diseases — disorder (entropy > 0.95 or order < 0.3) and stagnation — through REORGANIZATION_PRIORITY = "stagnation_first" (switchable to disorder_first / all), because a dissipative structure presupposes a flow: with zero flow no self-organisation can happen, while a disordered system still carries noise that structure can capture. The substantive correction: a single snapshot **cannot** separate "efficiently stable" from "dead" — all-success gives sigma 2.0 and entropy 0, and a genuinely frozen system looks identical — so stagnation now demands window evidence via an optional streak argument, standing only when streak >= 5 **and** entropy <= 0.10; with no streak the output is stagnation_side = None (not evaluated, not faked), and sigma < sigma_min is demoted to a weak_flow_signal instead of a verdict. Legacy needs_reorganization keeps its disorder-only meaning, with needs_reorganization_any added for downstream callers. (3) emergence_chain now **consumes** consistency_gate and entropy_production through dissipation_view / consistency_view / advisories, while its boolean judgement keeps exactly one source of truth (quartet_tensor / P2xP3 >= 0.7, P1 >= 0.6, P4 <= 0.4, order > 0.30): wiring the gate and the entropy production into the verdict would open a second and third truth source, and one disagreement would veto every emergence call. (4) the dify-workflow contract moves from "read" towards "run": a local loopback (our own http.server, not a Dify instance — there is none on this machine: five ports probed false, no docker, no app token) posts the real build_request payload over real HTTP and checks 15/15 — path /v1/workflows/run, Bearer header, the exact three-key body, blocking by default, the 11 response fields, 400 invalid_param when user is missing, and the streaming switch — so its verification now reads source_code_read+local_contract_loopback, paired with an explicit list of what that does not prove (Dify-side availability, real workflow output, real quota behaviour). New test_v67_contract_and_arbitration.py locks the 8 function key white-lists as exact sets (the user said 7; the measured consumption is 8 — rff_features is the V66 single-point sibling of rff_kernel), the 13-key / 10-key Kalman row contracts, the arbitration branches including a priority-switch test, and the static Dify contract: 43 cases; main suite 1476, redline 147, DOC raised to 0.9.61 -> **V66 resonance-pack-bodywork (the seven rulers move into the pack itself and the soft kernels are replaced — topology_resonance_pack.py now imports its math from resonance_supplement.py, so there is one source of truth and no drifting second copy; then: (1) mandala_fanout drops the _det_hash % 10 > 2 dice for a Hadamard transform over the 8 cube directions (round-trip error 0, orthogonality dev 0, Parseval ratio 1.0, DC equals the mean 4.5, and uniform loads leave only the DC term so effective fan-out drops to 0 — no divergence, no fan-out); (2) kalman_state rewrites the hardcoded single-source scalar filter into a predict step plus standard NIS (NIS = (z - x^-)^2 / (P^- + R), chi2(1) 95% = 3.841) with innovation gating: a deliberate outlier z=30 scores NIS 557.99 and is rejected, so the fused state stays 10.371429 instead of being dragged to 15.786207, and switching the gate off reproduces the legacy 15.786207 exactly — proving the old gap came from the missing predict step and gate, not from the data; kalman_fuse moves to the information form Y=1/P so q=0 makes sequential exactly equal to batch inverse-variance weighting (1e-9) and check() stays 7/7 PASS; (3) bayesian_guard replaces the hardcoded 0.5 threshold with the cost-sensitive Bayes-risk optimum C_fp/(C_fp+C_fn) = 1/11 = 0.090909 (network 0.6136 still BLOCK, read 0.0145 ALLOW) and adds an evidence path through Wald SPRT (A=2.890372 / B=-2.251292, stops at step 5); (4) hilbert_project replaces the fake projection (phi_k = 1/sqrt(k+1), v[k % n], norm diverging with dimension) with random Fourier features, so E[ |
| 上一批次 | V68 prehistory-ledger (the pointer ledger and the skeleton contract) — prehistory_ledger.py inherits the V1 list verbatim (590 top-level items, one authorised reclassification: the seven live sync/mirror facilities leave the prehistoric column) and prehistory_contract.py reads 22 runtimes by ast only (65 shared signatures across 16 files, asserts min 0 max 96 sum 1410, one registered zero-assert instance); 54 cases, main suite 1530, redline 149, DOC 0.9.62; previous — *V65 resonance-supplement (the seven bearer-only mechanisms stop being "implemented but mathematically soft" and gain real judges — resonance_supplement.py adds 9 literature anchors (Prigogine 1977 Nobel lecture on dissipative structures, the Royal Society 50th-anniversary review of minimum entropy production, Bak-Tang-Wiesenfeld 1987 SOC, Kalman 1960, NIS consistency testing, Wald 1945 SPRT, Rahimi-Recht 2007 random features, Kuramoto 1975) and a 12-entry upstream ledger (Kalman-and-Bayesian-Filters-in-Python 19480 stars / filterpy 3899 / TopologyLayer 431 / torchhd 394 / ripser.py 347 / javaplex-class TDA / sandpile 7 / Kuramoto 9 / stochastic resonance 19), all method_only or readonly_reference, no code copied; then 7 measurable judges replace the soft kernels: (1) C1/R8 entropy production sigma = sum J_i X_i with Onsager symmetry and Sylvester positivity (measured sigma=2.55, principal minors [2.0, 1.91], and sigma(tX)=t^2 sigma(X) holding to ratio 1.0 at t=0.5/2/3); (2) R2 mandala fan-out rebuilt as a Hadamard orthogonal transform over the 8 cube directions instead of a hash pretending success — round trip error 0, H8H8=8I, centre equals the mean (4.5 for loads 1..8), uniform input leaves only the DC term, Parseval ratio 1.0; (3) R5 multi-source fusion in information form P^-1 = sum R_i^-1 with heterogeneous R (fused 10.371429, P 0.190476, weights 0.190/0.762/0.048, NIS 0.116/0.038/0.078 all under chi2 3.841) plus sequential-equals-batch and a deliberate outlier case that gets flagged; (4) R6 guard becomes Wald SPRT with real thresholds A=2.890372 / B=-2.251292, stopping at step 5 on H1 and step 9 on H0 instead of a fixed window; (5) R7 Hilbert projection replaced by random Fourier features (Bochner) so the map is a real kernel approximation: MAE 0.198 at D=16 down to 0.035 at D=256 against the true Gaussian kernel, deterministic under a fixed seed; (6) X2 resonance gets an order parameter r= |
| Previous batch | V86 genome-DNA (three inner genes become fail-able criteria, and the meridian stops being a 10-fragment hardcode) — ①holoaeon/genome_dna.py verifies the three grounded genes: G-DNA-01 growth ≠ split (Δedges>0 ∧ Δcopies=0 ∧ shared protocol family ∧ shared seal → growth; Δcopies>0 → split), G-DNA-02 differentiation ≠ split (phenotype specialisation ∧ three-homology DIFF=0 across the 25 mechanism axes / heartbeat 70·768·17 / SEAL; θ never 1.0, theta=null before calibration, no "100% homology" claim), G-DNA-03 robustness in the real sense (R1 local failure metered, R2 under expansion pressure spec byte-identical + genome_hash DIFF=0 + cell count conserved, R3 three-homology conservation); ②meridian_link.py sheds six old skins — the len(rows)!=10 upper guard becomes a BASE_FRAGMENT_COUNT=10 lower bound merged with genome_dna.registered_fragments(), the table becomes dynamic, check_linkage uses expected = len(fragments) * len(KERNEL_MECHANISMS), and meridian_summary / build_ledger read len(load_fragments()) instead of the literal 10; ③new scripts/asset_census.py measured 616 top-level items (418 dir / 198 file) → 201 net candidates / 149 V3 family / 69 top-level zips / four-artefact-ready = 0 (the feed's "four pieces ready" premise does not hold on this machine — registered, not quietly redefined); ④new scripts/plugin_contract.py assembles 3 pilots under one eight-face contract and measures θ = 0.592 (min similarity over dual-state-true pilots; the third pilot reads 0.368 and is rejected as split/orphan — dual_state=false, never assembled); ⑤the registry appends 3 fragments, live meridian 300 cells (12 × 25), three PASS; ⑥the A14 always-true golden sentinel turned green by reading the DNA mother file's seal instead of constant-vs-constant (v3_gate 33/34 → 34/34), and the relative-import fix takes the redline scan to 159 files / 0 violations; ⑦DOC_VERSION not bumped — it lives in spec.py, and byte-identical there is the first criterion; ⑧the V85 governance batch is not executed (no technical dependency, the user asked sync-first) and is registered in honest_boundaries rather than dressed up. Main suite 2243 → 2285 · new test_genome_dna 42 cases · V62/V71 tests de-hardcoded (empty-extra_fragments baseline injection + single-source pf.EXPECTED_*) · manifest 341 entries unchanged (new files follow the V41-A tracking-only convention) |