gnomonlearn
Docs/Upstream formal repair and direct consumer evidence

Upstream formal repair and direct consumer evidence

Versioned draft snapshot

RFCs, reference catalogs, and evidence reflect Spec 3b2e3e7a114fb826b1445d39a48756b8fff629cd. Counts and pending/status notes belong to that snapshot. Read the dated project update for later progress.

Source repositories are private during staging. Links marked “repository access required” need access to the linked repository.

This is a trace and gap record from the pinned draft specification. It does not establish proof or implementation conformance.

Scope: actual upstream Lean API repair and three directly affected complete Gnomon consumer files. This stage is not full consumer adoption, publication, consistency certification, backend acceptance, or normative release readiness. The separate 13-area Gnomon umbrella check remains required.

Reviewable source commits

SourceLocal immutable candidateBase
Upstream fastverk/spec5ba3d99c182b138b4a150deebd6699a386c74286fetched main 8c777d196cbbd84dc48c9850a6d3a04dbc55fb64
Gnomon consumer preparation9bc57ee2c00ddf1a10c92072080bdbe788e402f2spec dff1e21acd3b05bb57282f46e340f95fa1a3e871

Independent writable checkouts are .work/formal-upstream-repair-20261013 and .work/formal-consumer-repair-20261013 under the workspace. The original upstream checkout/history remains untouched at 883f407. Canonical Gnomon still pins 3725bd9f6d2711aa9bb179deea84f51e4424382a; its known inconsistency blocker remains active. Delta bundles and mail patches preserve these local review commits. consumer-repair.* preserves the first consumer commit bd00174; consumer-repair-with-gate-fallback.* includes the subsequent fail-closed lexical fallback correction.

Actual checks

  • Matching native Lean: leanprover/lean4:v4.30.0-rc2.
  • All 18 tracked lean/Spec formal modules compiled from a fresh import tree. git ls-files '*.lean' also lists one Lake build-configuration file; it is not a formal theorem consumer. The 18 include Authoring, Compaction, Grounding, emitters, logic, admission controls, and FrameworkCheck.
  • //lean:spec_framework_test includes AdmissionCheck in dependency order; //lean:admission_test is a focused target. Compaction/Grounding remain separate BUILD targets and were included in the native source graph check.
  • Existing proof audit passed all 18: no source sorry/native_decide, all warning guards, trusted baseline 93 → 92. No new axioms were introduced.
  • Constructive actual controls prove invalid raw kind rejection, a nonempty admitted singleton, and rejection of conflicting joint membership. Named AX-016 and the corresponding controls have no axiom dependencies. AX-045 and singleton overlap control depend on the existing uninterpreted overlap symbol. This does not choose interval semantics.
  • Built-in #check_failure controls reject the old unguarded calls, missing membership, and missing overlap certificate. Lean prints recovered terms as informational diagnostics for these intentionally failing checks; they are not exported proof declarations. Actual inspected declarations have the dependencies reported above.
  • Official Mathlib cache tooling materialized 643 cached dependency archives at exact 5450b53e5ddc75d46418fabb605edbf36bd0beb6, matching the existing source manifest and Lean version. No stub or dropped import was used.
  • Algebra and Semiring compiled. Complete Foundation, Lifecycle, and Graph theorem files compiled after faithful Compat imports and explicit source names, with five authored deferred bodies. Native output retains all five sorry warnings; compilation is diagnostic-only.
  • Actual guarded consumer callsite controls reject absent membership or overlap evidence. Foundation/Lifecycle’s narrow uniqueness implications report no axiom dependencies, Graph’s implication depends on overlap, and all five deferred manifest declarations explicitly report sorryAx.
  • Focused synthetic gate controls passed: three original authored-gate tests, four historical known-pin tests, then four final authored-gate tests adding private/lemma/indented/attribute fallback cases. Synthetic compiler mocks are decision-path validation controls, not native Lean acceptance evidence.

Before/after and exact deferrals

baseline-source archives the exact original Gnomon input bytes used by the baseline commands. The first baseline harness failed on three missing old-pin Logic imports; consumer-baseline-missing-imports.log preserves that failure. After the actual old-pin closure was compiled separately, the original complete files had 61 body/type errors. The first repaired-API candidate had the same61 older rename/manifest failures. Final complete files have zero errors and five explicit deferred manifest bodies:

  • Foundation project_manifest_canonical_name
  • Foundation rfc0001_b33_grounded
  • Foundation rfc0002_b52_grounded
  • Graph rfc0040_b13_grounded
  • Graph rfc0040_b15_grounded

Their exact authored project.gnm propositions remain; external AX-034 still states project.aion. No branding axiom or false proof was substituted. The strict candidate gates separately count 718 generated deferrals and 5 authored deferred theorems, plus any unassigned code-level sorry tokens. Current unassigned count is 0. The lexical scan is conservative, not a Lean AST parser; unrecognized declarations remain blockers identified only by source location.

258 affected theorem-file hash bindings were refreshed using the existing review inventory; all 722 links remain reviewed with 715 semantic mismatches. The nonempty-kind finding now states local certification; guarded lifecycle and overlap findings retain their mismatch verdicts. These implications do not prove metadata preservation, write rejection/M36011, or backend admission.

Remaining publication/adoption work

Prepared PR body (repository access required) and publication sequence (repository access required) describe the breaking proof premises, remaining 92 assumptions, actual native scope, artifact/registry checks, license observation, and pending full consumer check. After an automatic-review rejection and explicit human approval, root published the exact upstream head as draft PR #81. This subtask performed no external write. No merge, tag, registry publication, canonical pin adoption, site generation/build, or deployment was performed.

evidence.textproto binds actual command logs/source/output bytes through the existing ValidationEvidence proto. verify.py verifies those bindings and focused outcomes; it does not run optional builds or certify consistency.

The full staged whitespace check reports original whitespace in preserved baseline sources, mail-patch context, and Git bundle-verification output. Those authoritative bytes remain unchanged. The authored Markdown/Python and typed evidence pass the scoped whitespace check.