RESEARCH — {427.1} OKF producer inversion: substrate index + delta measurements

RESEARCH — {427.1} OKF producer inversion

Task: id-427 — enumerate over the corpus, not over a type vocabulary. Spec-chain artefact: {427.1} RESEARCH (of {427.1}→{427.2}→{427.3}→PLAN). Date: 09/08/2026 (S546 spec wave). Posture: this is an index, not a re-measurement. The measured substrate already exists and is preserved point-in-time. This document names where each fact lives, records the delta measurements this pass executed that the substrate does not carry, and records the substrate interpretations it challenged.

---

1. The substrate — read this first, in this order

#	Document	What it carries
1	notes/s546-producer-current-state-audit.md	The measured current state: seven enumeration branches, four closed registries, the two holes verified by mechanism, the RunSummary finding, the wired-sites blast radius (§2), the index/ontology surfaces (§3/§4), review-set verdicts (§5–10), and its own coverage statement. Every file:line in this task starts here.
2	tasks/id-427.md	Goal, ACs, the S545 minting note and the S546 Progress note — the latter carries the three ratified spec directions and the two fold-ins. Those are owner rulings, not suggestions.
3	tasks/id-422.md	The scope_tag interlock: the publication gate that closes hole 1 by construction, and its deadlock guard (the gate lands with a write path, never before it).
4	../id-426-okf-v02-upgrade/notes/s546-okf-v02-spec-delta.md	The v0.2 emission this task’s output must be shaped for: generated:{by,at}, sources[], [^id] footnotes, no # Citations trailer, no top-level resource: for DB-backed concepts, okf_version: "0.2". In flight now; treated as landed target state.

2. Ratified inputs (DR-104: these outrank the code)

DR	What it binds here
DR-141	This task’s decision record. Enumeration corpus-complete by construction; type labels, never gates. The load-bearing consequence is the negative answer.
DR-019 (as amended S545/S546)	Divergence 1 — the closed validator-enforced type taxonomy — withdrawn. Version target v0.2. The CONFORMANCE.md house rule is re-opened (id-431).
DR-027 (as amended S546)	ontology.json reduces to the client-overlay carrier; the bundle-shipped base snapshot retires. The platform-repo half is unchanged. Implementation rides this task’s wave.
DR-050	The ingestion content_type axis is a different vocabulary. It is never merged into concept types, in either direction.
DR-054 / DR-079	The client-overlay contract and the four bundle classes. The overlay class gate (only client_business may compose a client overlay) is a live, separate requirement — it survives this task untouched.
DR-140	Entity resolution sharpens the entity-keyed grains; it does not close the coverage hole. Orthogonal, and not a dependency.
DR-139 / DR-123	An invariant, a task directive, a spec, or a DR’s own ratifying text is evidence of intent at that time, never of correctness. Schema and functionality change is on the table.
DR-060	A drafting-config or output-shape change is a manual version= bump on enrich_concept, recorded in the bundle log.md — never an automatic invalidation. This task triggers it (see TECH §5).
DR-016 / DR-025 / DR-047	Client ownership; the knowledge-admission (promotion) gate as the published-only read posture; the narrowly-scoped degrade posture (failed rather than abort).

Ratified product framing: initiatives/core-product/knowledge-base-foundations/okf-platform/bundle-doctrine.md (id-71 affordances A17/A19/A21, four bundle classes, two production paths).

3. Interlocks — what this task must not decide for someone else

Task	Interlock
id-422	Ratified: the residual grain is designed for hole 2; hole 1 closes by construction under id-422’s publication gate and is enumerated defensively until that gate lands. This task does not touch scope_tag’s write path or the CHECK constraint.
id-426	Owns the v0.2 frontmatter migration. This task emits into that shape and must not re-litigate sources[]/generated. Sequencing: id-426 first, or the two coordinate on frontmatter.py.
id-428	Owns the trust slot (confidence → verified/status). This task makes confidence: no-content reachable for the first time; id-428 maps it to status: draft. Named coupling, not a dependency.
id-429	Owns the index axis. This task fixes the directory axis it will index (TECH §2.4) — the coupling is named explicitly for the id-429 designer.
id-430	Closed (research). Its ruling — ontology.json overlay-only — is executed here.
id-431	Owns CONFORMANCE.md’s fate. This task withdraws superset 1 only and leaves the file’s existence to id-431.
id-358	Re-scoped to a rename executed in this task’s rework. The BI-28 slot stays.
id-362 F1	Closed as a named design input here: unify enrich_concept over the Source protocol, not over l_records’s concrete types.
id-163	Paused. This task changes gate mechanics only on system_baseline; it does not re-author that lane.

---

4. Delta measurements executed by this pass

Everything below was read at the cited file:line in /Users/liamj/Documents/development/canonical at main d1e6e14ad during this pass. These are not in the substrate.

#	Measurement	Why it is load-bearing
M1	type is producer-authored, never model-authored: enrich.py:939 passes type=key.concept_type into build_concept_frontmatter. The model’s Pass-1 envelope supplies title/description/tags/body/citations + the three hints — never type.	Opening the vocabulary does not hand type-minting to the model. The label stays deterministic and memo-stable.
M2	policy, capability, metric and dataset are emitted by no code path anywhere — they exist only inside RECOGNISED_FACET_TAGS (validator.py:147-149). The only facet tag actually emitted is reference (web_pass.py:169, applied :674). playbook is emitted only as a type, by the system-baseline lane (validator.py:246; repo_docs.py:18,30).	”Resolve the three bends as types” is, for policy/capability, a registry deletion, not an emission change. Only reference has a live emitter to re-point.
M3	canonical_facet_tag / normalise_facet_tags (validator.py:160-175) have no production caller — the only references repo-wide are scripts/tests/test_producer_validator.py:449-493.	The alias-fold mechanism is test-only. Retiring it moves no production behaviour.
M4	CONCEPT_TYPE_VALUES (lib/ontology/concept-schema.ts:70) has no runtime consumer — only its own module docstring and __tests__/lib/ontology/concept-schema.test.ts:73. The live colour legend is a separate array, lib/okf/concept-type-tokens.ts:38-48 KNOWN_TYPES.	The “TS legend” registry is a duplicate legend with no reader; the legend duty is already discharged elsewhere.
M5	Bundle directories are grain constants, not type derivations: hard-coded literals at l_records.py:845 (topics/), :864 (products/), :891 (company/overview.md), :924 (certifications/), :949+:982 (case-studies/). Only the feeder grain derives a directory from a type string (:815, f"{concept_type}/…"), and its own docstring already rules that an arbitrary type name has no principled English-plural rule.	The layout needs no type→path function. See C2.
M6	q_a_pairs.source_document_id is nullable and is NULL for derived_from_form_response pairs (supabase/migrations/20260621105625_id59_qa_pairs_source_document_id.sql:5-6,24). q_a_pairs.source_form_instance_id exists and is the won-bid lineage column (l_records.py:511,517).	A published pair may have no parent document. The residual attribution is a three-step cascade, not a single document key.
M7	_REPO_DOCS_BUNDLE_CLASSES (flow_def.py:138, consumed :519) selects which Source class a run instantiates (bundle-doctrine’s two production paths). It contains no concept type.	Substrate §2 lists it under “must change”. Requirement-first it is untouched by DR-141 — see C1.
M8	derive_concept_confidence (frontmatter.py:167-193) returns only strong/partial; _CONFIDENCE_VALUES (frontmatter.py:158) admits no-content/needs-SME, currently unreachable from any producer path.	The residual grain gives no-content its first real producer — the honest signal for a concept with nothing distilled in it.
M9	RunSummary (bundle_writer.py:528-570) carries no denominator: write_bundle derives removed as previous_paths − written − moved − failed (:1351), and every other field from drafts that already exist (:1325-1348).	Confirms the substrate: there is no field a corpus item could appear in without first becoming a ConceptKey. The fix must add the denominator, not another failure bucket.
M10	iri_projection._DIMENSIONS (:76-80) classifies base-vs-overlay per dimension using ALLOWED_CONCEPT_TYPES / ALLOWED_ENTITY_TYPES / ALLOWED_RELATIONSHIP_TYPES; project_context (:227-245) drops every overlay term when client_id is None.	Retiring ALLOWED_CONCEPT_TYPES removes the classifier for one of the three projected dimensions. context.jsonld cannot be left to fall through.
M11	The showcase bundle holds 19 concept .md files across the five directories plus case-studies/won-bid/ (find … -name '*.md' minus the three reserved files). There is no references/ directory — Pass-2 has minted no reference concept into the shipped bundle.	Migration surface is exactly these 19 files; the reference-as-type change has no shipped artefact to migrate.
M12	write_bundle raises on a physical write-path collision before any write in the run (bundle_writer.py:1310-1320).	The residual grain’s rel_paths must be collision-free by construction, not by luck.
M13	_won_bid_case_study_redirect / bundle_write_path / bundle_write_path_for_key (bundle_writer.py:224-284) exist solely because two case_study grains were forced to share one directory.	Once a grain owns its directory, the whole identity≠physical-path mechanism dissolves.
M14	repo_docs.py:117-119 computes SYSTEM_BASELINE_CONCEPT_TYPES from EffectiveOntology.base_for_class("system_baseline"), and RepoConceptKey.__post_init__:192-201 gates on it. repo_docs.py:234-242 declares its own duplicate Source protocol.	The fourth registry has a downstream in the second Source; id-362 F1’s protocol unification has a second, unrecorded duplicate to collapse.

5. Substrate interpretations this pass challenged

The substrate is measured; its interpretations are not automatically right (DR-104 applies to ratified docs, not to an agent report). Four were checked.

• C1 — flow_def.py:519-533 is listed as “must change”. Challenged and reversed. Requirement-first: it serves source selection (which Source a bundle class runs — bundle-doctrine’s Path 1 / Path 2), current source bundle-doctrine.md §“Two production paths”. DR-141 says nothing about it. Disposition: KEPT, with the rename ripple only.

• C2 — “the type-derived directory layout” (substrate §1). Challenged and corrected. Measured (M5), the layout is grain-derived; type and directory coincide today only because each grain emits exactly one type. This is the single most consequential correction in this document: the open-vocabulary design needs no type→directory function, and the existing five directories need no migration.

• C3 — “the facet-tag vocabulary … as a materialisation” to be amended. Challenged. The registry states its own two requirements (validator.py:120-133): (a) name the facets the producer treats as first-class for downstream consumers, and (b) keep the enumerated type set and the facet vocabulary from silently diverging. (b)‘s subject stops existing under DR-141. (a) is discharged today only by the TS colour legend (M4). Neither requirement survives in Python. Disposition: deletion, not amendment — while facet tags themselves survive as free-form tags: entries, which never needed a registry (check_concept has never rejected an unregistered tag).

• C4 — DR-141’s “every published unit lands in exactly one concept”. Challenged on measurement. A published pair carrying a scope_tag and a pattern-matched parent document already lands in two concepts today: the topic grain clusters it by tag, and the product/case_study/certification reads pull it via _SQL_QA_BY_SOURCE_DOCS_OR_ENTITY (l_records.py:377-382, source_document_id = ANY($1)). Partition (=1) is not the current behaviour and is not what the product needs; coverage (≥1) is. TECH §2.1 decides for coverage and recommends a DR-141 rider recording it.

6. What this pass did not cover

• No database was queried and no producer run was executed. Every claim is static.
• Test bodies unread beyond the four assertions grepped for M3 (test_producer_validator.py:449-493) and M4. The substrate’s test-file list is filenames only; assertion counts remain unswept.
• The vendored SPEC was not re-read this pass — id-426’s S546 delta note is relied on for every v0.2 claim.
• .claude/worktrees/agent-* stale producer copies unswept (still open from the substrate).
• specs/id-132-* DR-141 propagation unchecked (still open from the substrate).
• parse-index.ts / parse-log.ts internals and the /okf render path unswept — the census additions to log.md (TECH §5) touch a surface whose parser this pass did not read. Flagged in TECH §7 as a subtask-level check, not an assumption.
• No live context.jsonld instance exists to diff against (the showcase publish predates {132.44}), so the F6 projection change is verified against the writer only.

No repo file was edited and no database write was performed by this pass.