gnomonlearn
Docs/Authored theorem citation review

Authored theorem citation review

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.

The 55 Lean RFC catalogs carry 722 provenBy links on normative statements. They name 721 distinct authored theorems in area-level Theorems.lean files; one name is shared by two RFC-0007 statements. The local declaration inventory resolves every link. That check establishes a name and source file, not that the theorem states the linked RFC obligation.

python3 tools/check_proven_by_reviews.py now inventories these links and checks source-bound semantic reviews in lean/proven-by-reviews.textproto. The ledger currently has 722 reviews: seven formal matches that rely on external axioms or local definitions, 715 mismatches, and zero unreviewed links. RFC-0036’s two external-ID citations have mismatch verdicts: their global AX-010 propositions omit the authored kind scope, and the no-reassignment theorem does not model deletion. Their formal repair still awaits the identity-scope decision. RFC-0049’s formal tenant filter and principal-relative aggregate match their narrow claims. RFC-0050’s write-guard and unrestricted-capability theorems match narrow formal definitions, without proving runtime enforcement. Most other cited theorems state generic axioms that omit the linked RFC’s concrete operation, data shape, diagnostic, authorization, or lifecycle requirement. CI validates the ledger and theorem references; the normative release gate also requires each link to have a matching review. The review binds the exact statement record, authored RFC file, and theorem source file by SHA-256. Proof compilation and runtime implementation evidence remain separate gates.

RFC-0020’s 2026-10-02 revision scoped delivery to the subscription’s effective read scope and made queue overflow an explicit diagnostic exception. The four source-bound reviews for b8, b11, b13, and b21 were rechecked against that revision. Their cited theorems still do not model overflow diagnostics, query-kind authorization, subscription-open scope checks, or the agent permissions block, respectively; all four remain MISMATCHES_CLAIM. Their narrative hashes were refreshed without changing verdicts.

The RFC-0006 wrapper and RFC-0007 status-layer source edits were reconciled against 51 existing proof-link reviews. Their theorem declarations did not change; RFC-0007 statement b75 now explicitly concerns a gnomon’s inferred epistemic status. Its cited theorem still does not establish the verification rule, so its mismatch verdict remains. The ledger again passes source-hash validation with 714 mismatches and no unreviewed links.

The RFC-0037/0038 legacy-conformance corrections and RFC-0040/0047/0048 session-contract changes were compared with their existing theorem findings. The cited theorem files did not change, and the revised narrative still exceeds what those generic theorems establish. Narrative hashes were refreshed for 51 affected reviews; RFC-0048 b9 also changed its statement record without changing its scope-widening claim. RFC-0048 b5 and b7 became foundational axioms with no provenBy citation, so their obsolete reviews were removed. The two axiom decisions remain visible in the catalog; the ledger now covers exactly the 722 live citations.

This is distinct from the 723 entries in Derivations/Proofs.lean: those entries are generated from per-RFC fragments and have their own review ledger. The two catalogs must not be conflated merely because their claims overlap. Reviewed mismatches include RFC-0036’s bitwise permission comparison and no-silent-dedup claims. Their authored theorems prove reflexive equality and generic boundary acceptance/rejection, respectively. Other mismatches omit required diagnostic behavior, Statement scope, or query parameters. RFC-0049’s empty-projection theorem proves algebraic identities but not the compiler optimization cited by its provenBy link. The authored theorem for RFC-0036’s (kind, externalId) uniqueness claim now has a mismatch review: it applies external AX-010, which treats an external ID as globally permanent, while RFC-0036 and the Graph backend use a kind-scoped key. See EXTERNAL_ID_SCOPE_PROPOSAL.md.

Until a link is source-reviewed, provenBy is a historical citation to a Lean declaration, not evidence that the cited theorem proves the RFC claim. No public proof guarantee should be inferred from it. Once the owner resolves the external-ID scope and other formal contracts, review each authored theorem against the exact RFC statement, replace mismatched theorem statements or withdraw misleading links, compile the resulting formal sources, and record the assumptions behind each matching proof.

The proposed publication and migration policy is in PROVEN_BY_PUBLICATION_PROPOSAL.md. It requires a decision before changing public RDF triples.

RFC-0036 b33 review correction (2026-10-13)

The earlier matching review of rfc0036_resource_statement_disjoint_ids was too broad. Pinned AX-011 compares tagged EntityId.resource and EntityId.statement values; their constructor distinction permits equal raw numeric IDs. A compiled observation constructs default Resource and Statement values with rawId = 0 and unequal tagged identities. That result does not prove RFC-0036’s exclusion of shared bigint IDs. The citation and theorem remain unchanged, and the source-bound b33 review now records MISMATCHES_CLAIM. The generated derivation remains deferred. No additional currently matching citation references the related Foundation tagged-ID theorem. The current ledger has seven matches, 715 mismatches, and zero unreviewed links; the release gate rejects the mismatches.

This correction accompanies the separate exact write-time XOR derivation repair described in PROOF_GAPS.md. Its conditional proof under AX-008 does not resolve numeric-ID allocation, pending identity decisions, or backend conformance.

External input consistency limit (2026-10-13)

The unchanged pinned framework derives False from its nonempty-kind premise and default Resource value, as the compiled source-bound observation records. The seven matching citation reviews classify proposition-to-claim scope; they cannot establish proof soundness under inconsistent imported inputs. Both strict proof and normative release gates now explicitly reject the known pin. No external model change or speculative consistency certification is made for other pins.

Approved narrative revalidation (2026-10-16)

The approved source changes in commit f4d98a8f4bb9a1b6fda15444c14a426db615e5a1 were rechecked against the existing citation findings: RFC-0003 makes source priority optional and assigns effective priority 0 before ordering and tie checks; RFC-0004 corrects two Authoring Shape cross-references from §10 to §11. Only the narrative fingerprints of 18 RFC-0003 and 25 RFC-0004 reviews were updated. Their cited statement records, theorem names, theorem files, verdicts, findings, and row order are unchanged.

All 43 reviews remain MISMATCHES_CLAIM. In particular, RFC-0003 b17’s generic boundary theorem still does not model effective priorities, tie detection, or M03210. The ledger retains 722 reviews, seven matches, 715 mismatches, and zero unreviewed links. This scoped revalidation establishes current source bindings; it adds no proof or implementation evidence. The known framework-pin consistency blocker and the proof and normative release holds remain unchanged.