Formal proof status
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.
Known consistency blocker in the pinned external inputs: the minimal
Falseobservation fromkindNonempty (default : Resource)compiles withoutsorryor new axioms. The default Resource has an empty kind, while the universal premise forbids it. Proofs importing this framework cannot establish sound normative guarantees. The compiled implication and proposition-review counts below remain inventories; strict proof and release gates explicitly reject the affected immutable pin. See compiled evidence.
The Gnomon-owned lean/Gnomon/Derivations/Proofs.lean was originally copied byte for byte from the full-history gnomon/lean/lean/Gnomon/Derivations/Proofs.lean checkout on 2026-10-01. Both files then had SHA-256 f5d4f11baca6ff273206c0707852b83e2e7f5c032a32b8bd2a4e07c56ff02fb1. The old copy remains in gnomon until its consumers move to a released spec artifact. The inherited kg.lean.GenerateProofs is an LLM pass tied to Aion paths and IRIs, so it cannot deterministically regenerate this Gnomon file. The inherited theorem bodies are maintained in 55 Gnomon-owned lean/Gnomon/Derivations/Parts/RFC-NNNN.leanfrag source files. The earlier recorded concatenated body had SHA-256 86bf0875f2bd8a0c5cd49c900855dfdae0558cae4921753195ea69a81c6b9bab. The RFC-0036 statement-endpoint heading was synchronized with its revised authored and Lean claim; its theorem body was not changed. python3 tools/emit_derivations.py --write emits the catalog with an honest header and computed inventory footer; without --write it is a CI byte-drift gate. This establishes reproducible assembly of inherited candidate bodies, not proof soundness or semantic conformance.
python3 tools/check_derivations.py --framework-root /path/to/fastverk-spec compiles the external formal dependencies and this Gnomon proof file with the pinned Lean toolchain. The inherited generated file initially failed with 79 Lean errors: 32 GnomonType/AionType, 22 GnomonFn/AionFn, 13 undefined bindsGnomon, 9 project.gnm/project.aion, 2 gnomonGrammar/aionGrammar, and 1 undefined Gnomon namespace. lean/Gnomon/Derivations/Compat.lean now supplies four definitionally equal names for the inherited formal vocabulary; the emitted catalog imports that file directly, and the checker compiles the checked-in source without staging a modified copy. Against external commit 3725bd9f6d2711aa9bb179deea84f51e4424382a, this reduced the failure to 10 errors, all involving the project.gnm versus project.aion assertion. Those ten incompatible bodies are now explicit deferred obligations in their per-RFC source fragments. At that checkpoint, the Gnomon file compiled with --allow-deferred: 723 theorem declarations, 721 entries containing sorry (477 inherited or earlier-deferred, 234 semantic-review deferrals, and ten formal-manifest incompatibilities), and 2 without sorry. The generated footer reports the actual counts without labeling those candidates as proved. This compile-only result does not resolve the formal contract conflict or establish independent proofs. The project filename conflict needs the owner’s public-contract decision. CI runs the strict gate and remains red while deferred obligations remain.
tools/check_derivation_inventory.py compares theorem-entry headings to the current Lean-emitted RDF derivations. It now accounts for all 722 current derivations: 721 entries in the generated proof file and RFC-0049’s authored structural theorem, linked by provenBy in its Lean RFC source. The gate checks that the external theorem declaration and RFC link exist. This is a metadata inventory check, not a soundness proof; CI runs it after uploading the preview artifact and before compiling derivations. The local deterministic emitter replaces the old Aion-oriented LLM generator as the maintained assembly path. New or revised proof bodies still require explicit formal and semantic review; the emitter does not synthesize them.
The RFC-0049 source statement b4 cites rfc0049_subgraph_identities, which exists in lean/Gnomon/Spec/Algebra/Theorems.lean and proves six identities for union, inter, and sdiff with the empty projection. The RDF derivation b4_via_finset_semilattice is present in the current Lean-emitted Turtle. The authored proof has no entry in the separate generated Derivations/Proofs.lean catalog; the inventory gate accounts for it explicitly. The six algebraic equalities establish that these particular rewrites preserve projection results; they do not establish that a compiler performs the RFC’s SHOULD optimization. Generator repair must retain this distinction and must not turn the behavioral recommendation into a claimed proven compiler property.
PROOF_COVERAGE.md is a generated per-RFC inventory from the current emitted
RDF and Lean source. Its drift gate counts both := by sorry placeholders and
standalone sorry tactics, reporting 718 deferred entries and 3 bodies
without sorry in the generated file. It separately counts RFC-0049’s linked
authored structural theorem, so all 722 current RDF derivations have an
inventory entry and zero are reported missing. The report does not imply that a no-sorry body encodes the
RFC claim faithfully or compiles in the current strict Gnomon build.
lean/semantic-reviews.textproto is the source-hashed ledger for current no-sorry
derivation candidates. python3 tools/check_proof_reviews.py checks each review’s
proof-fragment, authored-RFC, and Lean-catalog hashes, verdict, and finding;
--release requires every candidate to
be reviewed and matching. Two hundred thirty-two bodies were found to prove
weaker propositions than their RFC headings: forty-six individual findings and
186 bodies across thirty-six repeated-proposition groups. Each now has an opaque obligation and
sorry in its per-RFC source fragment, with its exact mismatch finding
preserved in PROOF_DEFERRALS.md. This removes those
bodies from the no-sorry candidate set; it does not prove the named claims.
The current ledger reviews three no-sorry candidates. Two RFC-0036
reviews match their formal claims: Resource.kind is nonempty and is treated as
opaque text. Their proofs depend on the external kindNonempty and
kindOpaqueStipulated assumptions, so they do not establish backend
conformance. Two other RFC-0036 bodies were reviewed and deferred because
their old propositions invoked global external-ID uniqueness from AX-010
without a kind key, and one did not model deletion. Their findings are
preserved in PROOF_DEFERRALS.md; formal resolution remains pending the
scope decision in EXTERNAL_ID_SCOPE_PROPOSAL.md.
The separate RFC-0049 authored structural theorem remains outside this
candidate ledger. The third reviewed candidate is the write-time predicate XOR implication
described below. Strict proof release still fails on the 718 deferred bodies.
The generated proposition reuse report groups exact proposition text across distinct RFC claims. It makes repeated generic propositions visible for triage; reuse alone is not a semantic verdict.
The separate authored theorem citation review inventories
all 722 provenBy links in the Lean RFC catalogs. Every named theorem exists;
all 722 links have source-bound semantic reviews, including 715 mismatches and seven
formal matches that rely on external assumptions or local definitions. CI
checks the link ledger, and the normative release gate requires all links to
match their claims. A Lean declaration name or a compiled theorem
does not discharge this semantic obligation.
The six hand-written area theorem files formerly described RFC-wide coverage
as “100% kernel-verified.” Their headers now identify candidate or structural
theorems and point readers to the semantic review boundary. Only Lean comments
changed; all noncomment lines compare byte for byte with the preceding
revision. Because the review ledger fingerprints whole theorem files, its
477 affected file hashes were refreshed without changing any verdict or
finding. Both review gates pass with the same 716 mismatches and two reviewed
no-sorry derivation candidates. This is a correction to proof-status claims,
not an increase in proved RFC coverage.
FORMAL_MANIFEST_PROPOSAL.md and its unapplied patch give the owner an exact
project.gnm formal-model change to review. In a disposable copy of the
pinned framework, that patch compiled all 723 generated theorem entries with
zero Lean errors using the existing compatibility import; that prior run tested
477 deferred entries. That prior source had 721 deferred entries: 477 inherited or earlier-deferred bodies, 234 semantic-review deferrals, and ten manifest-name incompatibilities. The approved RFC-0048 session rewrite subsequently removed two obsolete deferred claims, leaving 719. The pinned external sources compiled that Gnomon catalog with --allow-deferred. The RFC-0049 structural
theorem is tracked separately by the inventory gate. No external formal source or dependency
pin was changed by this test.
The external Spec.Kernel and Spec.Axioms sources compile at that pinned commit. They are external formal inputs, not files owned in this extracted repository. Their comments contain historical claims about which K pillars are axioms; the checked local declaration audit in tools/audit_kernel.py is the evidence for current forms. Compiling an external theorem or carrying a provenBy citation does not establish that a Gnomon normative statement is proved or implemented.
The pinned fastverk/spec checkout has no top-level license file, so the right
to copy its formal sources into this repository has not been established by
the local source audit. Preserve the pinned external dependency until source
ownership and redistribution terms are confirmed; copying files to satisfy a
path check would not resolve the formal contract mismatches above.
The standalone Lean corpus gate now emits generated/turtle/axioms.ttl from the 57-entry local axiom metadata catalog and generated/turtle/inference-rules.ttl from the 5-entry local rule metadata catalog. tools/check_grounding.py confirms that all 722 emitted nonaxiomatic theorem-like statements reach an axiom through derivations that name catalog rules. This is RDF graph grounding; the external kernel sources, generated Gnomon derivation proofs, and deferred cases above remain open proof obligations.
Owner: spec, with the external fastverk/spec framework maintainers for type compatibility. The per-RFC fragment assembly and generator drift gate are now present. Remaining acceptance: resolve all Lean compilation errors by editing source fragments or the formal contract, not the generated catalog; replace deferred cases with sound proofs or mark the corresponding claims unproved; pass the strict proof gate; map every published proof claim to the exact compiled theorem and assumptions. Keep public proof guarantees withheld until these conditions pass.
Current derivation inventory repair (2026-10-13)
The approved RFC-0012 fixture contract had reached its authored text, Lean
catalog, and emitted RDF, but its deferred derivation heading still named the
older AST-only requirement. The per-RFC fragment heading and deferral finding
now include successful canonical-AST fixtures and failing diagnostic fixtures
with rejection. The existing opaque proposition, theorem identifier, and
sorry body are unchanged. The generated catalog was regenerated from that
fragment; no theorem became proved.
The inventory gate now requires one document matching each catalog filename,
one nonempty literal for each required text field, one valid modality identity,
and one inference-rule identity for each derivation. It rejects ambiguous
scalar values, detached sections/statements/derivations, and linked nodes
without their emitted type. Previously Graph.value could choose among
multiple values and str(None) could disguise an absent field. Ten mutation
tests cover these failures and preserve the valid axiom-without-derivation
case. CI runs the tests, and the portable release gate now runs the inventory
check directly.
Focused validation passes for 722 current RDF derivations: 721 generated entries plus RFC-0049’s separately linked authored structural theorem. The regenerated coverage report has no missing or orphaned entries. Both source-bound proof review gates pass their integrity checks; the authored citation ledger still has 714 semantic mismatches. The strict proof gate still has 719 deferred bodies, and this repair does not establish compiler conformance, proof soundness, or normative release readiness. Lean compilation was not rerun for this comment-only proof-source update.
RFC-0036 write-time XOR repair and numeric identity limit (2026-10-13)
deriv_6a569d7c360a now states the exact write-time XOR requirement already
represented by RFC-0036 statement b25 and its reviewed authored theorem:
writtenWithPredicateId s ∨ writtenWithPath s, together with exclusion of
both forms. Its proof projects those two conjuncts from the unchanged pinned
AX-008. AX-008 depends on the external k2 boundary-fidelity assumption and
opaque write-presence predicates; this is a conditional formal implication,
not a proof of backend validation, missing-field resolution, or diagnostics.
The statement, derivation, and theorem identifiers are unchanged.
The proposed second repair was withheld after inspecting AX-011. It proves
EntityId.resource r.rawId ≠ EntityId.statement s.rawId by constructor tags;
it does not exclude equal numeric rawId values. The compiled validation
control exhibits default Resource and Statement values with equal rawId = 0
and distinct tagged IDs. RFC-0036 exposes both IDs as bigint and prohibits
sharing an ID. The b33 provenBy review is therefore corrected to
MISMATCHES_CLAIM, and deriv_f95533f9b55d remains opaque and deferred with
this precise finding. No numeric allocation assumption or identity policy was
added. The related Foundation resource_statement_id_disjoint theorem has
the same tagged proposition; no current matching citation review references it.
The deterministic emitter, both source-bound review checks, derivation
inventory, coverage and reuse reports pass. The pinned allow-deferred gate
compiles all 721 generated theorem entries and the focused controls using the
same compiled inputs. Deferred entries decrease from 719 to 718; reviewed
no-sorry candidates increase from two to three. The authored citation ledger
has 715 mismatches and seven matches across 722 links. Its release mode still
rejects publication. Persistent terminal evidence
records the scope, unchanged external pin, source hashes, and initial source
formatting failure. The overall normative release blockers remain open.
Known upstream consistency blocker (2026-10-13)
The actual Lean observation imports only the unchanged pinned Spec.Axioms:
theorem pinned_framework_derives_false : False :=
Spec.kindNonempty (default : Spec.Resource) rfl
It compiles, and #print axioms reports only Spec.kindNonempty. The external
Resource permits kind = "" and derives Inhabited; kindOf projects that
field. Its universal kindNonempty premise therefore contradicts an available
structural value. This is an input-model defect, not a backend counterexample
or authority to change Gnomon’s public Resource or identity contracts.
A shared known-blocker registry is keyed to the observed immutable framework
commit. The strict derivation gate refuses this pin even with zero deferred
bodies; --allow-deferred retains diagnostic-only compilation and reports the
blocker explicitly. The release gate reports it alongside existing gaps.
An unlisted pin is not certified consistent. Focused validation checks those
conditions and both consumers; the actual compiled observation supplies the
source-backed finding.
Upstream repair acceptance: represent valid Resources through a refinement or an explicitly guarded validity domain, eliminate the contradictory universal premise, and compile source-backed validity/contradiction controls plus all actual consumers. Review and adopt a new immutable pin through the existing formal dependency process. No upstream source, axiom, model, or pin changes are included here. The earlier clean ed9 generation is preserved as historical catalog-generation evidence and is not a soundness or publication result.
Unadopted validity and population repair candidate
The formal model admission proposal
preserves a second actual same-pin False derivation: unrestricted AX-016
equates two constructible Statements with equal SPO/time keys but distinct raw
IDs, independently of the nonempty-kind premise. A Resource-only repair is
insufficient. The independent Lean candidate uses local valid-value
certificates and a shared admitted population with explicit uniqueness
obligations; it has an inhabited witness and rejects joint admission of the
conflicting pair. It imports only Lean core and selects no interval, create,
wire-ID, or upstream ownership policy. The proposal, candidate, and actual
observations are review artifacts, not an adopted model, consistency
certification for the full framework, new dependency pin, or release acceptance.
The existing immutable known-pin blocker and all deferred/mismatch counts stay
unchanged.