Formal model admission repair 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: complete review candidate, unadopted. Owners: the external
fastverk/spec formal-model maintainers and Gnomon spec, with Graph Platform
and Storage reviewing the connection to actual backend populations. The
external model, immutable pin, public RFCs, wire identities, and runtime
policies are unchanged. This is one proposal for the existing consistency
blocker, pending owner selection before actual adoption.
The executable candidate is
proposals/FormalModelAdmissionCandidate.lean (repository access required).
It imports only Lean Init, declares no axioms, contains no sorry, and does
not import the inconsistent external Spec modules. It is deliberately a
small boundary model rather than a replacement for the entire graph model.
1. Two source-backed contradictions to address together
The affected external source is the existing immutable pin
3725bd9f6d2711aa9bb179deea84f51e4424382a.
| Finding | Exact source boundary | Actual observation |
|---|---|---|
| Nonempty Resource kind | Universe.lean makes Resource.kind an unrestricted String and derives Inhabited; Predicates.lean defines kindOf as its projection and universally assumes kindNonempty. | kindNonempty (default : Resource) rfl derives False, without sorry or a new axiom. Existing compiled evidence. |
| Statement uniqueness | Universe.lean exposes a structural Statement constructor. K1/AX-016 quantifies over every such value, without a shared population or admission hypotheses. | Two Statements with nonempty-kind Resources and equal SPO/time fields but raw IDs 0 and 1 are equated by AX-016. Projecting rawId derives 0 = 1, hence False. Preserved actual observation. |
The second observation does not depend on kindNonempty or sorryAx; the
printed dependencies include Spec.k1 and its predicate symbols. Resource
refinement alone therefore cannot repair these inputs. Conversely, population
membership alone does not make every raw Resource’s kind nonempty.
The AX-016 native Lean subprocess exited 0. Its outer inline Python wrapper
later exited 1 because a post-compilation assertion incorrectly expected
only [Spec.k1] in the printed dependency list. The original source, native
log, version log, and typed evidence are preserved byte for byte. The wrapper
outcome is recorded separately as historical transcript metadata, not a native
Lean failure or a fabricated saved driver. It was not recompiled to hide that
wrapper error.
AX-045 also quantifies over all structural Statements. Its overlap predicate
is uninterpreted in the pinned source, so this review does not claim a
separate AX-045 contradiction or choose concrete interval semantics. Its
population scope still requires review before it can represent a backend
collection invariant.
2. Recommended boundary for review
Separate three kinds of values and evidence:
- Raw records remain constructible. They can represent invalid inputs or two individually well-formed rows with a conflicting key. Construction is not evidence that a value belongs to an actual graph.
- Local valid-value certificates justify properties of one value.
ValidResourcecarries a proof that its raw kind is nonempty. A statement’s local certificate checks its three Resource values and a caller-supplied temporal-validity predicate. These certificates alone do not establish cross-row uniqueness. - One admitted population carries its member list, local certificates for those members, and an explicit AX-016 uniqueness certificate over pairs in that same population. The uniqueness field is an obligation supplied with this population, not a universal axiom saying all raw lists satisfy it.
The candidate’s SameAX016Key mirrors the pinned proposition’s Resource
selectors and validFrom. Mapping those selectors to actual backend ID keys
is an explicit adoption obligation. Its Nat ID/time fields are demonstration
tokens; they do not select a wire encoding, Snowflake allocation, nullable
bounds, or identity policy. The candidate does not change equality to ignore
raw IDs or derive an ID from a triple/time tuple.
A separate OverlapCertificate population overlaps requires both rows to
belong to the same population. The overlap relation is supplied by the caller.
No half-open/closed, infinity/null, adjacency, interval-comparison, or create/
upsert rule is defined by this certificate. A proof conditional on that
certificate does not establish that it is obtainable for a real backend.
An arbitrary Workspace is insufficient: the current external Workspace has unrestricted membership predicates and an all-member default. Calling a raw collection a Workspace does not supply any admission invariant. A proposed adapter must obtain and preserve an actual certificate rather than invent a membership witness for every constructible value. The candidate list is a finite witness vehicle; its certificate uses membership, not list order or multiplicity. It selects no backend collection or deduplication representation.
3. Executable candidate and nonvacuous checks
The candidate provides:
certified_kind_nonempty, which obtains nonemptiness from the valid-value certificate, andempty_kind_has_no_certificate, which rejects an empty raw kind without prohibiting construction of that raw input;admitted_statement_uniqueness, which requires one population and explicit membership for each row;admitted_triple_nonoverlap, conditional on same-population membership and a caller-provided overlap certificate;- an explicitly inhabited valid Resource and a nonempty admitted singleton;
- two distinct raw Statements with equal AX-016 keys and separate local
certificates, plus
duplicate_pair_not_jointly_admitted; - another admitted singleton for the second row. The invariant does not prohibit raw construction or compare rows from different populations.
The duplicate-pair result rejects joint membership under admission; it does
not implement or select a write rejection API, diagnostic code, transaction,
allocation rule, or concurrency strategy. No insert/create transition is
provided. The local temporal predicate used by the witness is True only to
keep this example independent of pending interval decisions; it is not a
proposed interval-validity rule.
The six inspected proof declarations compile with no axiom dependencies. That establishes these small constructive consequences and an inhabited example. It does not certify a repaired external framework, full Gnomon model, compiler, or backend. Source, compiler, terminal output, and candidate hashes are bound in the candidate evidence.
4. Concrete signature impacts on three known theorem consumers
Actual sources remain unchanged. The following are the required shape changes if this boundary is selected; the executable candidate supplies the guarded proof pattern.
| Consumer | Current signature/body | Required reviewed signature and remaining meaning |
|---|---|---|
Foundation.Theorems.statement_uniqueness | Two unrestricted Spec.Statement values plus equal SPO/validFrom; applies ax_AX_016. | Add a shared admitted population and both membership proofs. Build SameAX016Key from the four equalities, then use the population’s uniqueness certificate. The result concerns admitted rows in that population, not all raw constructions. |
Lifecycle.Theorems.rfc0005_policy_affecting_preserved | The same unrestricted uniqueness implication is labeled as preservation of policyAffecting. | Its narrow uniqueness use needs the same population and membership arguments. This does not repair its semantic mismatch: real flag preservation needs authoring metadata, install-plan values, and a preservation relation across the relevant lifecycle stage. Its existing MISMATCHES_CLAIM review remains. |
Graph.Theorems.rfc0036_no_overlapping_triple_intervals | Two unrestricted Statements, distinctness and equal SPO; applies ax_AX_045. | Add shared population membership and an explicit overlap certificate/relation. The candidate proves only the guarded relation. The authored write-rejection and M36011 guarantee still require an operational/diagnostic model and backend evidence; the existing mismatch review remains. |
In Lean, the changed uniqueness use is concretely:
admitted_statement_uniqueness population first second
firstMember secondMember sameKey
The changed overlap use is concretely:
admitted_triple_nonoverlap population overlaps certificate first second
firstMember secondMember different sameTriple
The external Universe, selectors in Predicates, K1/AX-016 and AX-045
fields, named Axioms projections, defaults, and source-hashed reviews would
also change. Existing AX-001/AX-011/AX-019 structural proofs need recompilation
against any changed carriers. AX-011’s tagged-ID distinction still does not
prove disjoint raw bigint allocation; this proposal does not repair that
separate mismatch or select an allocation authority.
The local AX-016/AX-045 metadata catalogs, affected RFC statements and claim projections, generated references, provenance, and consumer pins must describe the same scopes after an actual adoption. Adding a compatibility alias cannot supply missing membership evidence or turn an invalid population into an admitted one.
5. Source ownership and upstream options
Recommended option for review: repair the external framework’s value and population boundaries upstream, retaining Gnomon as a pinned consumer. Its maintainers own the source change and release; Gnomon spec owns the RFC mapping, consumer proof signatures, review ledgers, generation gates, and adoption of a new immutable pin. Graph Platform/Storage own the link to persisted populations and runtime preservation. This option changes formal APIs and proof premises; it is not a branding rename.
A smaller upstream interface can retain raw carriers and use guarded validity
and admission hypotheses instead of exported valid-value wrappers. It must
remove or replace the contradictory unrestricted propositions. Keeping
those imports and adding guarded Gnomon wrappers leaves False derivable.
Independent per-row certificates cannot replace the shared population witness.
A Gnomon-owned independent model is another owner option. It would need a reviewed source/license boundary, complete formal ownership, explicit mappings, and consumer cutover. The present candidate is independently written and imports only Lean core; it does not vendor external sources whose redistribution rights are not established here or establish that Gnomon should take over their ownership.
An abstract carrier with hidden constructors is a third option, but it must provide an inhabited model of its selectors and constraints. Hiding a contradiction behind new opaque assumptions, or changing equality/identity to make the observation disappear, is not an accepted repair. No option is adopted by this proposal.
6. Exact acceptance before actual adoption
- Record the selected source owner and validity/admission boundary, preserving the pending statement identity/create decision, validity-bound decision, and ID allocation decision. State explicitly which propositions concern raw values, locally certified values, and one admitted population.
- Remove the contradictory unrestricted nonempty-kind premise and unrestricted AX-016 population quantification in the actual selected source. Review AX-045’s same-population scope; do not choose interval semantics through an unreviewed helper. Supply an inhabited admitted model, reject invalid local values and conflicting joint membership, and retain constructible raw input examples. A Resource-only repair is insufficient.
- Compile source-backed controls against that actual model: the two preserved
old-pin
Falseconstructions must no longer type-check through the repaired interfaces; valid-value and admitted-population witnesses and duplicate-pair rejection must compile. Inspect dependencies and permit nosorry, hidden universal replacement axiom, or inconsistent old imports in those controls. These checks are necessary and do not certify arbitrary full-model consistency. - Update and compile all three known consumers and every affected generated and authored theorem against the actual changed interfaces. Test that omitting validity or either same-population membership witness fails at those actual call sites. Review the complete external consumers, not just this toy model.
- Model successful mutation as preservation of admitted-population invariants and rejected mutation as a separate boundary/diagnostic outcome. Specify and verify real temporal/overlap behavior only after the existing contract choice. Test scalar/batch/concurrent writes and lifecycle/plan preservation in actual backends before claiming runtime guarantees.
- Regenerate formal catalogs, RFC references and RDF from source; rebind exact affected proof/projection reviews without promoting existing semantic mismatches. Build the complete external framework and Gnomon consumers, record exact sources/results, and review the new immutable dependency pin. The known old-pin blocker stays registered; another pin is not automatically certified consistent. Publish a coordinated pin/site update only within the existing owner-approved release workflow.
7. Current effect and limits
This proposal adds an independent executable review artifact and preserves the
second actual contradiction. It adopts no model, adds no production axiom,
changes no existing theorem signature or RFC/RDF identity, selects no runtime
policy, updates no external pin, and performs no site rebuild or deployment.
The strict known-pin blocker remains in force; the current 718 deferred bodies
and 715 mismatched authored citations are unchanged. The already verified site
at spec 42b93440e remains an explicitly blocked preview.