Curated Category-G Walk — Emission, Coverage Gate, and ODR-0024 Remediation

Context and Problem Statement

ADR-0031 planned the curated Category-G walk; ADR-0030 is the emission pipeline + gates (G1–G3). Executing the walk surfaced four engineering problems this record settles:

1. Emission must be emitter-driven, not hand-edited. The 188 candidate-G leaves are realised as generator code, regenerated to the committed TTLs; ci-byte-identity forbids hand-editing.
2. The walk needed an HONEST progress/completion metric. The existing G3 ci-descriptive-roundtrip gate measures whether the emitted SHACL profiles round-trip their form-question leaves — a property of the BASPI5 overlay (the one non-thin profile), not candidate-G TBox emission coverage. Treating it as the walk-completion test conflated the two and over-claimed “the curated walk has landed” while only 7/188 leaves were emitted.
3. A cross-module duplicate class slipped past CI. opda:riskIndicator was once declared in two module TTLs with conflicting rdfs:domain, passing ci-byte-identity + ci-three-graph undetected.
4. Council Session 028 (ODR-0024) returned corrections (monetary false-collapse, an unwarranted facility-class citation, a categoriser under-count) that must be reflected in the emitted model, or the model and the ratified ODR disagree.

Decision Drivers

• Byte-deterministic, emitter-driven emission (no hand-edited TTL).
• An un-gameable progress metric — the over-claim proved a profile-round-trip gate is a weak proxy for TBox coverage.
• No silent omissions and no false coverage (the ADR-0031 totality assertion; ODR-0022 §5 residue discipline).
• The council-ratified modelling rules (ODR-0024) must be the emitted model.
• Every opda: term declared in exactly one module TTL.

Considered Options

• A — emitter-driven walk + a dedicated candidate-G coverage gate + a dup-declaration gate + apply the ODR-0024 remediation (chosen).
• B — hand-edit the module TTLs directly — rejected: ci-byte-identity forbids it and it loses deterministic regeneration.
• C — treat the G3 ci-descriptive-roundtrip gate as the walk-completion test — rejected: it measures one emitted SHACL profile (BASPI5), not candidate-G TBox coverage; conflating them over-claimed completion at 7/188.
• D — defer the ODR-0024 remediation — rejected: leaves the emitted model and the ratified ODR in contradiction (the monetary false-collapse, the facility classes minted under a citation that does not license them).

Decision Outcome

Chosen “A”, because it makes walk progress an objective number the model is checked against, closes the cross-module-duplicate hole, and keeps the emitted model in step with the council-ratified rules — at the cost of one local-only gate and a one-time re-open of part of the committed walk.

The walk is realised as table-driven emitter code per semantic family (flat datatype/object properties per ODR-0024 R1), regenerated via opda-gen emit. Two CI gates are added to the ADR-0030 pipeline:

• ci-dup-declaration — fails if any opda: term is the subject of a defining rdf:type in more than one module TTL (the riskIndicator regression). Wired into the GitHub workflow.
• ci-category-g-coverage — reports X/N candidate-G leaves emitted-or-collapsed (N = the categoriser’s candidate-G count, 239 after R5; minted = local-name match; collapsed = registered in inputs/category_g_curation with an emitted target; uncovered = reported by name). Derives the candidate-G set from the categoriser (reproducible); local-only — it needs the gitignored data dictionary, so it reports UNAVAILABLE on a CI checkout. --strict gates at totality.

The ODR-0024 remediation is applied: R3 (withdraw the 18 monetary collapses → defer; honest coverage 179/239 after R5’s re-count), R4 (re-warrant opda:NearbyFacility off §Q4a onto the UFO Substance-Kind basis, collapse the subkinds into the genus, schoolType→SKOS), R5 (structural C-vs-G categoriser rule + re-run + regression test — a report-scope fix, no byte-identity re-pin), R6 (mint ~5 SKOS schemes from live enums), R7 (opda:AttachedDocument), R8 (titleNumber→RegisteredTitle), R11 (SHACL on mediaUrl/url + hasSubAssessment guard).

Consequences

• Good, because walk progress is an honest, un-gameable number (179/239 after the R5 re-count), and the model is emitter-driven + byte-deterministic.
• Good, because ci-dup-declaration closes the cross-module-duplicate class that byte-identity + three-graph missed.
• Good, because the coverage gate distinguishes a report-scope miscount (the categoriser under-count) from an emitted-artefact defect, so a count fix does not force re-pinning correct emissions.
• Bad, because ci-category-g-coverage is local-only (gitignored data dictionary) — a dev-time tracker, not enforceable in GitHub CI.
• Bad, because the ODR-0024 remediation re-opens part of the just-committed walk (monetary defer, facility re-warrant) — a one-time cost paid for council-correctness.
• Neutral, because the monetary leaves (pending opda:MonetaryAmount) and the categoriser-surfaced additional G leaves are routed to follow-on chunks, not emitted here.

Confirmation

Six gates gate the walk and run in pytest: ci-byte-identity, ci-three-graph, ci-profile-contract, ci-dup-declaration, ci-descriptive-roundtrip (BASPI5 profile round-trip), and ci-category-g-coverage (candidate-G TBox coverage; --strict at totality). Emission regenerates via opda-gen emit; a hand-edited TTL fails ci-byte-identity. Coverage settling at the honest 179/239 (the R3 de-collapse + R5’s candidate-G re-count 188→239) is the expected post-remediation state, not a regression.

Rules

• The two coverage gates are orthogonal and neither substitutes for the other. ci-descriptive-roundtrip = SHACL profile round-trip (a property of the emitted profiles); ci-category-g-coverage = candidate-G TBox emission coverage. Conflating them was the over-claim this ADR exists to prevent.
• Coverage definition: a candidate-G leaf is covered iff minted (its local name is declared in a module TTL) OR collapsed (recorded in inputs/category_g_curation.COLLAPSED with an emitted target). A collapse whose target is not emitted is a broken disposition the gate flags. Uncovered leaves are reported by name — nothing silently omitted.
• Emitter discipline: no hand-edited TTL; every minted term declared in exactly one module + listed in that module’s catalogue tuple.

More Information

• Modelling rules realised (depends-on): ODR-0024 — the Council Session 028 dispositions this engineering applies; ODR-0022 — the Category-G strategy + C/D/G boundary; ODR-0008 — §Q4a/§Q5a/§Q6a binding rules.
• Realises (implements): ADR-0030 — the emission pipeline these two gates extend; ADR-0031 — the walk execution plan.
• Deliberation: Council session-028 — the per-question verdicts + DA scorecard behind the remediation.