Imported from Blackcat-Informatics/gmeow-ontology (.agents/skills/gmeow-ontology-authoring/SKILL.md). Install upstream with npx skills add Blackcat-Informatics/gmeow-ontology --skill gmeow-ontology-authoring. Copyright stays with the author.

GMEOW Ontology Authoring & Verification

This skill guides agents in modifying, compiling, and validating GMEOW ontology resources in the current slices-first, Rust-native repository.

Guidelines

1. Constitutional Alignment:
   • Every design decision must align with CONSTITUTION.md.
   • Cite Constitution Principles (e.g. "Principle 4") in your pull requests and commits.
2. One Canonical Source (Principle 4):
   • Slices: Author ontology terms in slices/<group>/<name>/module.ttl. Slice metadata lives in manifest.ttl; slice-local docs, tests, examples, mappings, and translations live beside the module.
   • Mappings and correspondences: Author pure term linkage in the owning slice's mappings/equivalences.ttl. Author slice-owned projection cells beside that slice, normally as mappings/projections-<profile>.ttl. Shared or cross-slice projection enrichment may live in dsl/mappings/projections/*.ttl; shared transform function declarations live in dsl/mappings/transforms.fno.ttl.
   • Statements: Author statement-level metadata in dsl/statements/.
   • References: Author citations in metadata/references.ttl.
   • Generated artifacts: NEVER hand-edit generated/, root ontology-docs/, dist/, or generated mapping/projection/query outputs. Regenerate them from canonical sources.
3. No-Drift Gate (Principle 7):
   • Run make check after canonical source changes that affect generated artifacts: it materializes them through the single producer and then gates the result, so it is the whole no-drift proof in one pass.
   • Run make check before proposing, committing, or submitting changes unless the user explicitly narrows validation.
4. Logic projection doctrine (Principle 17):
   • Treat logic: as the canonical formal layer. OWL, SHACL, ShEx, SSSOM, EDOAL, FnO, SPARQL projection queries, and OWL alignment axioms are generated projection dialects, not sources of truth.
   • Existing hand-authored shape surfaces are transitional. Do not satisfy new modeling or slice-quality findings by adding hand-authored SHACL or ShEx unless the user explicitly scopes work to the legacy migration path.
   • A slice still shipping a hand-authored shapes.ttl (or contributing to a root shapes/*.ttl) carries per-slice migration debt, surfaced by the slice-quality.projection.ungrounded-shape advisory. Discharge it by re-expressing each shape as a logic:Constraint / OWL axiom in module.ttl that PROJECTS to generated/shapes/*, proving parity, then retiring the hand-authored file — the exact procedure in docs/MIGRATING-SHAPES-TO-LOGIC.md. Migrate under equivalence-before-deletion: never delete a shape whose check is not yet reproduced by the projected union.
   • Cross-ontology linkage and projection should be represented as correspondence work: use the current slice-local native alignment cell (a reified skos:*Match statement carrying gmeow:sssomFile/gmeow:justification/gmeow:confidence) and gmeow:ProjectionMapping frontend honestly so it can lower to logic:Correspondence with the right relation, direction, loss, law, and preservation claims.
5. Reasoner-driven flagship counter-examples (Principle 17/18):
   • When authoring or improving a slice's flagship acceptance scenarios (gmeow:FlagshipScenario), drive each guarding counter-example to the reasoner: the malformed input should make the native solver observe the missing or failed entailment at reasoning-runtime, raising the scenario's named failure class — not merely trip a structural/SHACL well-formedness shape. Because SHACL is a lossy projection of the canonical logic: core (Principle 17), a counter-example that bites only at the projection surface leaves the negative space unproven; run it through the native solver (Principle 18) so the failure is raised in the core.
   • The structural/SHACL proxy is the floor; the reasoner-driven counter-example is the depth target. Mark each scenario honestly with gmeow:counterExampleDischarge (gmeow:structuralDischarge today, gmeow:reasonerDrivenDischarge once the solver is actually run over the counter-example) — never assert reasoner-driven for a counter-example the solver is not run over. The flagship_counterexample_depth slice-quality axis reads that marker and surfaces which counter-examples remain structural-only.
   • See slices/grounding/logic/design/LOGIC-CONFORMANCE.md (Tests as ontology data) for the counter-example depth expectation.

Actionable Instructions

• Orient first: Read .goals, .baseline, AGENTS.md, the relevant slice docs, and the relevant slices/grounding/logic/design/ documents before editing ontology semantics.

• Edit ontology terms: Modify the owning slice, normally slices/<group>/<name>/module.ttl. Keep every GMEOW-namespaced term annotated with rdfs:label, skos:definition, and rdfs:isDefinedBy.

  make validate

• Edit translations and the per-slice glossary: A term's annotation coat and its fr/zh translations are one term-batch and land together — never add a coat and defer its translations. A new coat grows the localizable-literal denominator, so an unpaired coat dilutes measured translation coverage and reds make slice-quality-gate against the slice's raise-only axisTranslationCoverage floor; ratchet that floor to the freshly measured value on each landing. Translations live in slices/<group>/<name>/i18n/{fr,zh}.po; they are the only home for non-English ontology prose.

  The per-slice terminology glossary is a derived object — the pipeline folds every reviewed .po pair into a gmeow:Glossary in graph/lang-glossary-corpus, resident in gmeow.gts (never a hand-authored sidecar). make i18n-lint hard-rejects cross-batch terminology drift — one term translated two different ways across batches (lang:GlossaryTermInconsistency) — unless the source is an explicit lang:DeclaredTerminologyHomograph. That cross-node consistency check is authored as a logic:Constraint + logic:Formula in module.ttl (the functional-dependency dual of the GMN alias-injectivity bijection) and enforced by the Rust detector; never hand-author a sh:NodeShape for it.

  make i18n-lint

• Edit cross-ontology linkage and projections:

  1. Put pure identity or match linkage in the slice that owns the match subject, usually slices/<group>/<name>/mappings/equivalences.ttl.

  2. Put lossy projection legs, profile bindings, guards, transforms, and loss notes in the owning slice's mappings/projections-<profile>.ttl; use dsl/mappings/projections/*.ttl only for shared cross-slice enrichment.

  3. Choose the honest relation. Do not force equivalence where the mapping is an overlap, bridge view, lossy lens, or affine correspondence.

  4. Regenerate and check generated mapping/projection artifacts:

     make mappings
     make check-generated

  5. Validate Wikidata syntax and links when QIDs/PIDs are touched:

     make wikidata

• Edit statement provenance: Open and edit files in dsl/statements/, then regenerate and check drift:

  make check

• Assess slice quality: Use the slice-quality advisor at the start and end of slice uplift work:

  make slice-quality SLICE=slices/core/tags
  make slice-quality

  Interpret projection and linkage findings through the logic: / correspondence-calculus roadmap. For example, a missing shape surface points toward canonical logic constraints and derived validation shapes, not a new hand-authored shapes.ttl.

• Migrate the slice's hand-authored shapes (do this whenever you touch a slice): The Shape Migration slice-quality axis lists every hand-authored sh:NodeShape in the slice's shapes.ttl that carries no logic:formalizes (slice-quality.projection.ungrounded-shape). Each is a second source of truth to migrate into the canon. Prove and clear them:

  gmeow-dev slice-quality slices/<g>/<s>      # lists the ungrounded shapes
  gmeow-dev shape-equivalence --path slices/<g>/<s>   # EQUIV ⇒ the projector reproduces it

  Author each obligation in module.ttl with a reasoner-safe antecedent so the projector reproduces the shape, then delete the block — never owl:cardinality / owl:minCardinality / owl:maxCardinality (out of the EL fragment; they red make reason-verify and block make check):

  • at-most-one → a owl:FunctionalProperty (→ sh:maxCount 1);
  • existence → owl:someValuesFrom <Class> (→ sh:class; the dropped sh:minCount is a design-sanctioned ValidationOnly under-approximation);
  • class/datatype → owl:allValuesFrom; disjunction → owl:unionOf; faceted range → owl:onDatatype + owl:withRestrictions; cross-node check → a logic: FOL assertion in crates/pipeline/src/stages/constraint_shapes.rs.

  A genuine residue the fragment cannot express (exactly-N cardinality, node-level sh:or, bespoke cross-node sh:sparql) instead keeps its block but adds logic:formalizes naming its canonical logic: source — the form the blanket projection-purity gate legalizes. docs/SLICE_GUIDE.md §9 is the reference.

  The axis (axisShapeMigration) climbs the ladder Grounded 0.60 / Linked 0.75 / Exemplified 0.85 / Maximal 0.95, and each slice is floor-gated at its own committed score with monotonic non-regression — the gate forbids the measured score falling below the committed floor, it does not force ascent. The floor is ontology-resident: a gmeow:AxisFloorCommitment individual (gmeow:floorSlice / gmeow:floorAxis / gmeow:floorValue) authored in slices/core/slice-quality-rubric/module.ttl, projected read-only to generated/governance/slice-quality-axis-floors.tsv (never hand-edit the TSV — it is a generated projection). Raising a committed floor is a deliberate hand-edit of that individual (raise-only, hard-fail-enforced), made once the measured score has genuinely risen — never bump it ahead of a real uplift. To seed a not-yet-floored axis, emit its commitment at the live measured score and paste it into the rubric module.ttl:

  gmeow-dev slice-quality-seed-floors --axis axisShapeMigration   # or --all-axes

  The seeder is emit-only and one-shot per axis: it refuses to lower an already-committed floor, so re-running it never regresses a guarantee.

• Raise the slice's GMN-1 coverage (touch this whenever you extend a slice's vocabulary): The GMN-1 Coverage slice-quality axis (axisGmn1Coverage) reports the fraction of a slice's own GMN-0 normal-form vocabulary (module.ttl + examples/*.ttl) the GMN-1 codec (crates/lang-bridge/src/gmn1_codec.rs) can losslessly round-trip (slice-quality.gmn1-coverage.uncovered). It realizes gmeow:gmnCorrNormalToGmn's logic:mnemomorphic true claim, whose declared domain is ALL of GMN-0: the grounding slices (slices/grounding/{logic,lang,math}) are hard-gated at a committed floor of 1.0; every other slice is floor-gated at its own committed score — an ontology-resident gmeow:AxisFloorCommitment individual authored in slices/core/slice-quality-rubric/module.ttl (projected read-only to generated/governance/slice-quality-axis-floors.tsv), with monotonic non-regression — the gate forbids falling below the committed floor, it does not force ascent toward 1.0.

  gmeow-dev slice-quality slices/<g>/<s>      # names every uncovered GMN-0 quad
  gmeow-dev slice-quality-gate                # the per-axis floor gate (make slice-quality-gate)

  Round-trip each named construct by extending the codec's covered fragment, or file a named codec-coverage gap against LANG-GMN.md if it is genuinely out of scope. Raising this slice's committed floor — the gmeow:AxisFloorCommitment individual in the rubric slice's module.ttl (projected read-only to generated/governance/slice-quality-axis-floors.tsv), seedable at the live score with gmeow-dev slice-quality-seed-floors --axis axisGmn1Coverage — is a separate, deliberate hand-edit made once the measured score has genuinely risen — never bump the floor ahead of a real uplift.

• Run ontology reasoning and verification:

  make reason
  make verify
  make reason-verify

• Run full validation before handing off:

  make check