gnomonlearn
Docs/RFC-0001 module cardinality proposal

RFC-0001 module cardinality proposal

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.

Status: implemented and verified routine source alignment. Owner: spec / module authoring contract. Observed 2026-10-14 at spec commit 4734dc26c2b979ea9ae25ec1657895f8762e5f99.

Existing contract and error

Authored RFC-0001 §5 rule 2 requires exactly one module declaration and assigns M01060 to zero or more than one. The Lean diagnostic catalog gives the same condition; Lean normative statement b29 requires exactly one top-level module declaration and separately rejects other top-level declarations. Neither contract needs to change.

The Lean exactly-one-module-declaration invariant summary and the RDF domain summary instead say zero or more, which includes the valid one-declaration case. Before correction, the source-bound RDF review correctly reported this as one of 2,421 mismatched nodes among 6,954 reviewed nodes, with none unreviewed. The corrected node now matches; 2,420 mismatches remain. RFC-0001 retains all 138 projection reviews and 15 other mismatches.

The earlier RFC-0000 section-title abbreviations are already corrected in the current sources and are not part of this proposal.

Applied exact correction

Both source summaries replace:

module.gnm MUST contain exactly one module declaration; zero or more produces M01060.

with:

module.gnm MUST contain exactly one module declaration; zero or more than one produces M01060.

The two-line source patch (repository access required) changes only kg/source-domain/RFC-0001.textproto and lean/Gnomon/Spec/Foundation/Rfc0001.lean. The maintained correction keeps the domain IRI https://gnomon.sh/domain/invariant/exactly-one-module-declaration, all claim IRIs, classes, provenance, diagnostic code, and enforcement phase. The owning generators regenerated RDF, reference output, and the diagnostic source snapshot. All affected source-bound projection reviews were refreshed.

This changes a consumer-visible RDF literal and a generated reference summary. It restores the existing cardinality rule; it does not add a parser restriction or prove parser behavior.

Measured consumer effect

The local counterfactual loads the current ontology, claim and domain Turtle, and the RFC/formal catalogs retained by the accepted clean site build. It contains 89,301 triples. Replacing only the domain summary removes one triple and adds one; graph size and identities remain unchanged.

All 38 checked-in query files were evaluated before and after the edit. Thirty-seven result multisets are identical, including all 16 claim queries. governance/invariants-by-rfc.rq changes one row by replacing the incorrect summary with the corrected wording. Exact changed rows, query source hashes, and graph input hashes are in the typed counterfactual evidence (repository access required).

Exact-string searches of tracked files in the other five extracted repositories find the IRI and old wording only in gnomon’s inherited duplicate spec tree. The typed search results (repository access required) record each repository head, command, exit status, and output. Dynamic queries and future consumers are outside this search. The site consumes staged spec artifacts, so its reference page and invariant query should display the new wording after the corrected spec is pinned and the site is regenerated.

No deployment exists, so this requires no deployed-data migration. The framework duplicate and released consumer artifact still need cutover.

Observed implementation checks

The actual command record (repository access required) contains all twelve successful generation and integrity commands, their exact arguments, process IDs, exit statuses, and hashed logs. The pinned Lean corpus compiled all 55 structured modules covering 55 of 57 authored RFCs. Projection drift and semantic-review integrity, proof-review integrity, provenBy-review integrity, diagnostic drift, RDF links, and all 16 claim queries passed. These ordinary integrity gates preserve the separate strict-release failures: 715 provenBy mismatches remain, including all 25 RFC-0001 citation mismatches.

The actual delta verification (repository access required) checks the following requirements against regenerated output:

  1. The exact two-line source patch preserves all other authored RFC and Lean statements and the existing M01060 description.
  2. The reference invariant changes exactly once. All 57 generated Turtle catalogs, including axioms and inference rules, are byte-identical to the pre-correction catalogs. The source invariant is emitted into Markdown; the current RFC Turtle emitter does not emit that invariant summary as an RDF literal.
  3. Domain Turtle contains exactly one removed summary triple and one added summary triple, with no IRI change.
  4. All 138 RFC-0001 projection reviews bind the changed Lean source hash. Every unrelated finding and verdict is preserved. Only this node changes to MATCHES_SOURCES, leaving the global mismatch total at 2,420.
  5. All 742 diagnostic definitions and the diagnostic coverage report remain unchanged. The diagnostic snapshot changes only RFC-0001’s source hash.
  6. The repeated 38-query comparison against actual regenerated output has one corrected governance row and 37 unchanged result multisets, including all 16 claim-query result multisets.

Consumer cutover remains pending: pin the corrected spec artifact in the framework and site, verify the displayed summary, and remove the duplicate after its consumers move. Parser conformance for zero, one, and two module declarations remains a separate implementation gate.

Implementation authority and limits

The program goal authorizes resolving stale or conflicting source and generated facts. The root implementation review confirmed the existing unequivocal contract in authored §5 rule 2 and the M01060 diagnostic, then directed this routine correction. Selecting new syntax, identities, diagnostic behavior, or an ambiguous normative rule would require the separate owner decision specified by PROMPT.md; this correction selects no new contract. The owner has confirmed that nothing is deployed, so no deployed-data migration is involved.

Release, public publication, dependency-pin changes, parser conformance, and the corrected upstream formal model are outside this implementation’s authority and evidence.

The compact evidence record retains byte-identical copies of the original observations and relevant source, with an exact hash-bound original-to-copy manifest. Typed absolute paths still name the original execution inputs and outputs. The copied files are reviewable records rather than a standalone reproduction environment or final clean package. The genuine generation receipt predates the final maintained gap-doc update; final committed artifact generation must produce a fresh receipt.

The original counterfactual did not compile Lean or change canonical inputs. Its first reader copied the RDF graph, producing unrelated result changes from query iteration order; that reader and output are preserved under historical-copy-graph-comparison/ in the evidence directory. The corrected reader queries the same store before and after the single triple replacement, preserving the iteration order of unrelated triples.