diff --git a/DECISIONS.md b/DECISIONS.md index 48e128d02..a1629bf88 100644 --- a/DECISIONS.md +++ b/DECISIONS.md @@ -38,3 +38,7 @@ Append-only. Historical source outcomes remain in their original records. | 2026-11-01 14:40 | 2 | B/tooling | Registered the empirical statement check (main#726, merged d0d384b1; tool blob 020cb194) under GOVERNANCE and STATUS with its exact commands, the hosted navigation run 41242533228 / job 124659189973 (35 tests ok on CPython 3.11.17 at the merge commit) and the three-class mutation contract; coverage recorded honestly: main formal 31/31 PASS, #740 formal 59/59 PASS, Math- formal 63/63 UNSUPPORTED (exact-rational interval backend is the recorded gap). No global CLOSED guard installed; the final G0-G11 run remains required. Actor: Anthropic Claude session 01NMeK (tool author); reviews of record 01DKKe (same family, P1 fixed by follow-on) and Grok Bot agent 12 (different family). | Yes | [registration](GOVERNANCE.md#empirical-statement-check-for-core-lean-arithmetic-skeletons-bounded-registration); [#726](https://github.com/d6g8k5htny-coder/main/pull/726) | [2026-11-01 22:05 UTC] [Phase 3] [Layer C / G8] Correct the stale current-navigation statement that PR #713 is pending integration. Native PR metadata records merge 9abd2c100ae105103945073a7fa6c47b15679269 at 2026-10-31 10:18:44 UTC. Preserve the former 11e28815 source as pre-integration history, all scientific snapshot bytes and all qualification limits. This documentation correction neither replays checks nor establishes current-composition, theorem or CBZ-0 acceptance. Reversible: isolated commit; previous text retained in Git history. Evidence: https://github.com/d6g8k5htny-coder/main/pull/713 and https://github.com/d6g8k5htny-coder/main/commit/9abd2c100ae105103945073a7fa6c47b15679269 . + +| UTC | Phase | Layer | Action and rationale | Reversible | Evidence | +|---|---|---|---|---|---| +| 2026-11-06 05:30 | 2 | B/tooling | Amended the G12 registration (GOVERNANCE, STATUS) after main#742 (ℝ backend, sampled binders, CORROBORATED verdict), #743 (finite discrete models) and #752 (merged a7295bf2: census mode over every Lean declaration of Math- dfa9b30a; qualified-name resolution with declaration scope; every declaration a target; Codex findings 4/4 repaired; tool blob 6e7d6c1e, 86 tests): commands incl. `--lean-dir`, hosted navigation run 42165008899 / job 127719321978 (86 cases ok on CPython 3.11.18 at the merge commit); coverage recorded honestly: main formal 59/59 PASS, Math- formal 3·56·4, census 7·128·0·3,647 classified; row field `G12: PASS | CORROBORATED | not applicable ()` with the #695-index annotation (34·69·264·0); the exp/log interval gap narrowed, not withdrawn (no registered formal/ statement uses them; the census classifies 49 declarations as transcendental — 33 Real.exp, 9 Real.log, 4 Real.cos, 3 Real.sin, 39 of them P15 proofs — which carry `G12: not applicable (transcendental …)`), the other classes named. No global CLOSED guard; final G0-G11 run still required. Actor: Anthropic Claude session 01NMeK (tool author); declaration-scope repair by a successor lane; Root B-003 integration. | Yes | [registration](GOVERNANCE.md#empirical-statement-check-for-core-lean-arithmetic-skeletons-bounded-registration); [#752](https://github.com/d6g8k5htny-coder/main/pull/752); `cbz/evidence/g12-statement-fuzz-01n/README.md` | diff --git a/GOVERNANCE.md b/GOVERNANCE.md index 808079227..360107001 100644 --- a/GOVERNANCE.md +++ b/GOVERNANCE.md @@ -124,27 +124,42 @@ means to closure, never a reason to open a general infrastructure project. ### Empirical statement check for core-Lean arithmetic skeletons: bounded registration -**Registered 2026-11-01 14:40 UTC; global admission of the empirical layer remains pending.** -[main PR #726](https://github.com/d6g8k5htny-coder/main/pull/726), merged as -`d0d384b18a409ed51061458287dde22816782210`, supplies `tools/cbz/statement_fuzz.py` -(blob `020cb194c206d268972fcc9f50dd0707b21819d0`, SHA-256 -`1bd0f4a3059d6f040cb978691d2954deecf9a84287ce2272e48fe9ae93bdd7a7`), -`tests/test_cbz_statement_fuzz.py` (blob `723bbd306c70f0ea66b2c461ea40da9df4d923b6`) -and the run records under `cbz/evidence/g12-statement-fuzz-01n/`. - -Purpose: the addendum's empirical layer for *statements*. Kernel verification is not -strengthened by sampling, so for every theorem registered in a formal package's -`manifest.json` the tool parses the Lean statement and the package's own `def`s in an -exact-arithmetic interpreter for the core-Lean subset these skeletons use, evaluates it -independently of Lean (Python integers and `Fraction`; a float anywhere is an assertion -failure), enumerates finite quantifier domains exhaustively (list membership, or an upper -bound on an explicit `Nat`/`ℕ` binder only), probes just outside each domain for tightness -and vacuity as information, and reports PASS / FAIL / UNSUPPORTED per target. Lean core -semantics are implemented where they differ from naive arithmetic: `Nat` truncated -subtraction, totalized division (`x / 0 = 0`, `x % 0 = x`), `Nat` floor division, -Euclidean `Int` `/` and `%` (`Int.ediv`/`Int.emod`), numeral typing by the `binop%` -maximal-type rule. A FAIL is a misstatement or transcription defect in the Lean text, -never kernel evidence; UNSUPPORTED is fail-closed and makes the exit code non-zero. +**Registered 2026-11-01 14:40 UTC; amended 2026-11-06 after main PRs #742, #743 and #752; global +admission of the empirical layer remains pending.** +[main PR #726](https://github.com/d6g8k5htny-coder/main/pull/726) (merged +`d0d384b18a409ed51061458287dde22816782210`) supplied `tools/cbz/statement_fuzz.py`; +[#742](https://github.com/d6g8k5htny-coder/main/pull/742) (merged `d479620b`), +[#743](https://github.com/d6g8k5htny-coder/main/pull/743) (merged `fe327e02`) and +[#752](https://github.com/d6g8k5htny-coder/main/pull/752) (merged +`a7295bf20741c111da06d72faa846536ff76fa1c`) extended it. Current source at that merge: +`tools/cbz/statement_fuzz.py` (blob `6e7d6c1ecd489bcddf5f3b819ed6a9327fcda4a6`, SHA-256 +`78ec967a2a2957a203c57b412769d7685f9d953817b5173d4102a813c7120f1d`), +`tests/test_cbz_statement_fuzz.py` (blob `daab0995bdbd0e4779ba2c4b336a2eaf53a6c22b`, 86 tests) and the run +records under `cbz/evidence/g12-statement-fuzz-01n/`. History: the 2026-11-01 registration bound tool blob +`020cb194` (35 tests; hosted run 41242533228 / job 124659189973) with coverage main `formal/` 31/31 PASS at +`4712fc13`, #740's candidate `formal/` 59/59 PASS at `43b10c6f`, Math- `formal/` 63/63 UNSUPPORTED at +`19a55750`; those receipts stay under the evidence directory and in Git history. + +Purpose: the addendum's empirical layer for *statements*. Kernel verification is not strengthened by +sampling, so the tool reads each Lean statement and the package's own `def`s — by Lean-qualified name, +resolved from the statement's namespace outward and through its `open`s, with definition bodies evaluated +in their own declaration scope — in an exact interpreter: Python integers, `Fraction`, and the +multi-quadratic field ℚ(√d₁, …, √dₖ) for `Real.sqrt`; a float anywhere is an assertion failure. Lean-core +and Mathlib conventions are implemented where they differ from naive arithmetic: `Nat` truncated +subtraction, totalized division (`x / 0 = 0`, `x % 0 = x`), Euclidean `Int` `/` and `%`, `0⁻¹ = 0`, +`Real.sqrt x = 0` for `x ≤ 0`, exponents typed on their own (`rightact%`), ascriptions elaborated at the +ascribed type, numeral typing by the `binop%` maximal-type rule. Verdicts, fail-closed: **PASS** — a closed +statement (finite quantifier domains enumerated) evaluates true exactly; **CORROBORATED** — a statement +with binders before the colon (ℕ ℤ ℚ ℝ under decidable hypotheses; measure-theoretic statements on finite +discrete models Ω = `Fin n` with rational weights, tables and subsets) was refuted by no sampled instance +drawn from an exact pathological pool, the hypotheses' own boundary values and seeded rationals — admitted +evidence, never a proof and never PASS; **FAIL** — an exact witness, a misstatement or transcription defect +in the Lean text, never kernel evidence; **UNSUPPORTED**, naming the construct the interpreter does not model +(structures, transcendental functions, finite vectors `Fin n → ℝ`, unbounded quantifiers inside a +statement, `∑`, …). UNSUPPORTED is never upgraded and makes the exit code non-zero; `--admit-corroborated` +admits CORROBORATED into exit 0 and is recorded in the report. Census mode (`--lean-dir`) evaluates every +`theorem`/`lemma` of a directory without a manifest, accounts for every declaration (one the splitter cannot +separate is an UNSUPPORTED target, never an omission) and classifies the UNSUPPORTED reasons. From a checkout of this repository, run: @@ -152,50 +167,58 @@ From a checkout of this repository, run: python3 -B -S -m unittest tests.test_cbz_statement_fuzz -v python3 -B -O -S -m unittest tests.test_cbz_statement_fuzz -v python3 -B -S tools/cbz/statement_fuzz.py --package formal --selftest --json +python3 -B -S tools/cbz/statement_fuzz.py --lean-dir --exclude /raw --json ``` -Self-verification: the 35 tests (tokenizer, Lean semantics including the six Euclidean -division goldens of main#684/6511137449, definitions, folds, pairs, quantifiers, vacuity -and tightness probes, fail-closed behaviour, the mutation contract, the real package) ran -on the merge commit in the hosted `navigation` job -([run 41242533228](https://github.com/d6g8k5htny-coder/main/actions/runs/41242533228), -job `124659189973`, CPython 3.11.17, artifact `research-navigation-41242533228` whose -`commit.txt` is `d0d384b1…` and whose `tests-normal.log` shows the 35 -`test_cbz_statement_fuzz` cases `ok` among 1,892) and locally in both modes with -`-W error`. `--selftest` grades every module by a three-class mutation contract: -guaranteed-false mutants must FAIL (each comparison replaced by its negation one conjunct -at a time, the right-hand literal of an equality shifted by one — never `<`→`≤`), -truth-preserving weakenings (`<`→`≤`, `>`→`≥`) must PASS, and the manifest's declared -negative controls that change a statement must FAIL; an UNSUPPORTED or errored mutant is -a survivor, never a kill, and a stubbed evaluator fails the unmutated control. Reviews of -record: same-family 01DKKe -([726/6547986788](https://github.com/d6g8k5htny-coder/main/pull/726#issuecomment-6547986788); -one P1 fail-open domain inference found and fixed by its follow-on, tree `198c1f97`) and -different-family Grok Bot agent 12 -([726/6548050253](https://github.com/d6g8k5htny-coder/main/pull/726#issuecomment-6548050253), -PASS with the follow-on); Codex 01a103c8 supplied the division-semantics controls. - -Coverage actually observed, per package and cut: - -| Package | Cut | Targets | PASS | UNSUPPORTED | Mutation contract | -|---|---|---:|---:|---:|---| -| main `formal/` (SIDE24 skeleton) | `4712fc13` (`formal/` unchanged since) | 31 | 31 | 0 | 98/98 killed, 12 weakenings kept, 7 declared controls FAIL | -| main #740 `formal/` (SIDE24 + P15 skeletons, candidate) | `43b10c6f` | 59 | 59 | 0 | 184/184 killed, 14 declared controls FAIL | -| Math- `formal/` (`ResearchFormalCoreR1`, Mathlib real analysis) | `19a55750` | 63 | 0 | **63** | — | - -Gate mapping: G0/G5/G6 use the documented tests and the mutation contract; G8/G9/G11 -require this exact scope, runnable commands and self-verification record. For a new -CLOSED claim on a core-Lean arithmetic statement, a PASS at the same source cut as the -kernel receipt is the applicable empirical validation this registry recognises, recorded -with the report file's SHA-256 and the `manifest_sha256` it binds; it adds nothing to and -takes nothing from G2–G4. The tool does not cover Mathlib real analysis, binders before -the colon or unbounded quantifiers, so the Math- core's 63 statements remain explicitly -unsupported by the empirical layer; an exact-rational interval backend for -`Real.exp`/`log`/`sqrt` (offered by Claude017, main#684/6532742483) is the recorded gap. -This component installs no global CLOSED guard and does not complete the common final -G0–G11 run. Maintenance owner: none; stdlib only, self-contained. Actual executor: -Anthropic Claude, Claude Code session `session_01NMeKEismAyeqgdB4sy2NJU`; scientific -effect NONE. +Self-verification: the 86 tests (tokenizer; Lean semantics with the six Euclidean division goldens of +main#684/6511137449; exact real algebra; sampled binders and hypothesis boundaries; finite models; +qualified-name resolution and declaration scope; declaration splitting; census mode; fail-closed behaviour; +the mutation contract; the real packages) ran on the merge commit of #752 in the hosted `navigation` job +([run 42165008899](https://github.com/d6g8k5htny-coder/main/actions/runs/42165008899), job `127719321978`, +CPython 3.11.18, artifact `research-navigation-42165008899` whose `commit.txt` is `a7295bf2…` and whose +`tests-normal.log` shows the 86 `test_cbz_statement_fuzz` cases `ok` among 2,051) and locally in normal, +`-O` and `-W error` modes. `--selftest` grades every module by the three-class mutation contract: +guaranteed-false mutants must FAIL (each comparison replaced by its negation one conjunct at a time, the +right-hand literal of an equality shifted by one — never `<`→`≤`), truth-preserving weakenings (`<`→`≤`, +`>`→`≥`) must hold, and the manifest's declared negative controls that change a statement must FAIL; an +UNSUPPORTED or errored mutant is a survivor, never a kill, and a stubbed evaluator fails the unmutated +control. Mutants carry the binders, so a class-1 mutant of a sampled statement must be refuted by sampling. +Reviews of record: #726 — same-family 01DKKe +([726/6547986788](https://github.com/d6g8k5htny-coder/main/pull/726#issuecomment-6547986788); one P1 +fail-open domain inference, fixed by its follow-on) and different-family Grok Bot agent 12 +([726/6548050253](https://github.com/d6g8k5htny-coder/main/pull/726#issuecomment-6548050253)); Codex 01a103c8 +supplied the division-semantics controls. #742 — Grok Bot agent 8 (engineering, PASS_WITH_NITS at +`48f3a461`), agent 12 (integration), agent 10 (readback). #752 — Codex automated review in two rounds, four +findings (returned-function parameter types; definitions pooled by short name; declarations the splitter +missed; definition bodies evaluated in the caller's scope), all repaired at `95e5de31` and `55571918` and +answered in their threads; Root (OpenAI/Codex, B-003) exact-head qualification and integration +([752/6676699342](https://github.com/d6g8k5htny-coder/main/pull/752#issuecomment-6676699342)). + +Coverage actually observed, per package and cut (receipts under `cbz/evidence/g12-statement-fuzz-01n/`): + +| Package | Cut | Targets | PASS | CORROBORATED | FAIL | UNSUPPORTED | Mutation contract | +|---|---|---:|---:|---:|---:|---:|---| +| main `formal/` (SIDE24 + P15 skeletons) | `ec6962ff` (`formal/` unchanged since) | 59 | 59 | 0 | 0 | 0 | 184/184 killed, 24 weakenings kept, 15 declared controls FAIL | +| Math- `formal/` (`ResearchFormalCoreR1`, Mathlib real analysis) | `19a55750` = `dfa9b30a` | 63 | 3 | 56 | 0 | 4 (named measure-theoretic constructs) | 72/72 killed, 6 weakenings kept; receipts at seeds 1 and 2, seed 3 recorded in the evidence README | +| Math- every Lean declaration, 16 groups (census, `--lean-dir`) | `dfa9b30a` | 3,782 | 7 | 128 | 0 | 3,647 (classified by construct) | — | + +Gate mapping: G0/G5/G6 use the documented tests and the mutation contract; G8/G9/G11 require this exact +scope, runnable commands and self-verification record. For a CLOSED row on a Lean statement, the field this +registry recognises is `G12: PASS | CORROBORATED | not applicable ()`, recorded with the report +file's SHA-256 and the `manifest_sha256` (package) or source-file SHA-256 (census) it binds, at the same +source cut as the kernel receipt: PASS and CORROBORATED are statement-level evidence, *not applicable* +keeps the row's kernel status and names the construct, and a FAIL blocks the row until the Lean text is +repaired. A per-record annotation of the #695 literal kernel index on this rule is +`cbz/evidence/g12-statement-fuzz-01n/census_math_dfa9b30a_20261104/g12_annotations_literal_index_82d9e480.json` +(367 records: 34 PASS · 69 CORROBORATED · 264 not applicable · 0 FAIL). None of this adds to or takes from +G2–G4. The recorded gap (an exact-rational interval backend for `Real.exp`/`log`) is narrowed, not +withdrawn: no statement of the registered `formal/` packages uses those functions, but the census classifies +49 declarations as *transcendental function* (33 `Real.exp`, 9 `Real.log`, 4 `Real.cos`, 3 `Real.sin`; 39 of +them in the P15 proofs of `specifications/lean_statement_proofs_20261011`), which therefore carry +`G12: not applicable (transcendental …)`; the other UNSUPPORTED classes are named in the census. This +component installs no global CLOSED guard and does not complete the common final G0–G11 run. Maintenance owner: none; stdlib only, self-contained. Actual executor: Anthropic Claude, +Claude Code session `session_01NMeKEismAyeqgdB4sy2NJU`; the declaration-scope repair of #752 by a successor +lane under the owner's account; scientific effect NONE. ### Finite MomentTail empirical component: bounded registration diff --git a/STATUS.md b/STATUS.md index 1f52a8cf7..0bf545dd3 100644 --- a/STATUS.md +++ b/STATUS.md @@ -46,9 +46,11 @@ not full AST/type/theorem synchronization. New gating tools require documented commands and self-verification before admission; unavailable or unsupported coverage remains explicit. No documented command here runs all G0-G11 yet. -The empirical statement check for core-Lean arithmetic skeletons is a registered -bounded component on `main` ([PR #726](https://github.com/d6g8k5htny-coder/main/pull/726), -merged `d0d384b18a409ed51061458287dde22816782210`). From a checkout, run: +The empirical statement check (G12) is a registered bounded component on `main` +([PR #726](https://github.com/d6g8k5htny-coder/main/pull/726) `d0d384b1`, extended by +[#742](https://github.com/d6g8k5htny-coder/main/pull/742), [#743](https://github.com/d6g8k5htny-coder/main/pull/743) +and [#752](https://github.com/d6g8k5htny-coder/main/pull/752), merged `a7295bf20741c111da06d72faa846536ff76fa1c`). +From a checkout, run: ``` python3 -B -S -m unittest tests.test_cbz_statement_fuzz -v @@ -56,13 +58,17 @@ python3 -B -O -S -m unittest tests.test_cbz_statement_fuzz -v python3 -B -S tools/cbz/statement_fuzz.py --package formal --selftest --json report.json ``` -It evaluates every registered Lean statement of a formal package independently in exact -arithmetic with Lean-core semantics and grades itself by a mutation contract; the +It reads every registered Lean statement of a formal package (and, in census mode `--lean-dir`, every +declaration of a directory) independently of Lean in exact arithmetic with Lean's conventions, samples binders +before the colon at exact pathological values and on finite discrete models, grades itself by a mutation +contract, and reports PASS / CORROBORATED / FAIL / UNSUPPORTED fail-closed; the [registration](GOVERNANCE.md#empirical-statement-check-for-core-lean-arithmetic-skeletons-bounded-registration) -records commands, hosted self-verification and limits. At [registration #741](https://github.com/d6g8k5htny-coder/main/pull/741): main `formal/` 31/31 -PASS; the Math- core's 63 real-analysis statements are UNSUPPORTED, which is the honest -coverage at that source cut, not a pass. A PASS is statement-level evidence -only: not kernel evidence, not scientific acceptance, not the common G0-G11 run. +records commands, hosted self-verification, limits and the row field +`G12: PASS | CORROBORATED | not applicable ()`. At `a7295bf2`: main `formal/` 59/59 PASS; Math- +`formal/` 3 PASS · 56 CORROBORATED · 4 UNSUPPORTED (named); every Lean declaration of Math- at `dfa9b30a` +(3,782): 7 · 128 · 0 · 3,647, each UNSUPPORTED classified by construct. CORROBORATED is admitted evidence, +never a proof; a PASS is statement-level evidence only: not kernel evidence, not scientific acceptance, not +the common G0-G11 run. The finite MomentTail component is a registered bounded candidate in [Math PR #545](https://github.com/d6g8k5htny-coder/Math-/pull/545), exact source