Doctrine-curator lens — exemptions, accreditation, provenance

Squad report, 2026-07-26. Profile-loaded (doctrine-daphne), read-only. Evidence base for creed-and-values-design-hardened.md. Directives applied: DIRECTIVE_044 (canonical sources), DIRECTIVE_003 (decision documentation), DIRECTIVE_032 (conceptual alignment), DIRECTIVE_018 (doctrine versioning), DIRECTIVE_043 (close defect classes by construction).

Findings

[BLOCKER] creed-and-values-design-as-proposed.md:47-49 — "all behavioural doctrine artefacts (assets and templates are exempt)" diverges from the canonical exclusion set _NON_AUGMENTATION_ELIGIBLE_KINDS (src/doctrine/artifact_kinds.py:201-203), which excludes {TEMPLATE, ASSET, ANTI_PATTERN}. The design omits anti_pattern — the one kind structurally incapable of carrying the field. It also never defines "behavioural", leaving toolguide / mission_step_contract / glossary_pack unruled. Recommendation: express the exemption set as a named frozenset in artifact_kinds.py per that file's own docstring rule at :30-32 ("no second kind enumeration may be re-declared elsewhere").

[BLOCKER] src/doctrine/drg/models.py:292, :328 — DRGNode and DRGEdge have no model_config at all → Pydantic v2 default extra="ignore". Part 2's impacts field would be silently dropped on load. The DRGNode.tags docstring at :305-311 documents this exact hazard verbatim: tags had to be explicitly modelled "so an authored tags: [...] key round-trips instead of being silently dropped on load." Recommendation: add extra="forbid" to both before any edge-field work, with the round-trip assertion in the same commit.

[BLOCKER] src/doctrine/drg/migration/extractor.py:1210-1217, :1219-1229 — _node_to_dict and _edge_to_dict are explicit field-by-field writers (urn/kind/label/tags; source/target/relation/when/reason). Even a correctly modelled impacts vanishes at graph regeneration. Note DRGNode.provenance / DRGEdge.provenance are already deliberately omitted here, so graph-level provenance does not survive regeneration and cannot be relied on. Recommendation: model + writer + round-trip test in one commit, or do not add the field.

[BLOCKER] src/doctrine/agent_profiles/profile.py:254 — AgentProfile.model_config = ConfigDict(populate_by_name=True). No extra="forbid", no frozen. A value_bias: key loads and silently vanishes at model level; only the generated agent-profile.schema.yaml:7 additionalProperties: false catches it, and only where the schema validator actually runs. This is the single most likely way this design ships inert.

[HIGH] src/charter/schemas.py:286 — CharterYaml.model_config = ConfigDict(frozen=True, extra="forbid"). A creed: key at the root of .kittify/charter/charter.yaml is rejected today. Recommendation: add a typed creed: CreedConfig | None = None. Do not route it through the untyped overrides: dict[str, Any] escape hatch at :305 — that reproduces the unvalidated-slot problem. If the creed lives in a separate creed.yaml, it needs its own model and a place in charter/hasher.py + charter/versioning.py or it never participates in charter freshness. That is Part 3's real unstated gap.

[HIGH] creed-and-values-design-as-proposed.md:41-52 — one concept covers two semantics: a weight ("how important is this to us", :43-45) and a delta ("how does this support/weaken", :46-49). §5.1's own analysis says why they differ. Deltas compose by summation; weights do not. Recommendation: two fields, two Pydantic types.

[HIGH] src/doctrine/tactics/built-in/analysis/ammerse-impact-analysis.tactic.yaml — zero attribution: no notes block, no sddevelopment/Crossland provenance, no URL. Line 61 cites "the first-order impact matrix from the AMMERSE practice article" — an unresolvable external pointer with no toolguide. Both an unaccredited use of a trademarked system and an undocumented external reference. Close in the same change — it is the artifact a reader lands on when asking "what are these seven axes?", so it is where accreditation is actually read. Cheapest, highest-value item on the list.

[HIGH] LICENSE + repo root — LICENSE is MIT, "Copyright GitHub, Inc."; the repo does not assert its own authorship there. There is no NOTICE file. Appending an AMMERSE trademark line to this LICENSE would be actively misleading about who attributes what. Recommendation: create NOTICE at repo root; verify it is in the sdist/wheel packaged data or the accreditation does not ship to consumers — precisely the "so it is not lost" failure mode.

[HIGH] src/doctrine/anti_pattern.graph.yaml:5-33 — anti_pattern nodes carry exactly urn/kind/label/tags. There is no field able to hold a rationale, and the design makes rationale mandatory. An anti-pattern literally cannot satisfy the design's own invariant. Compounded by extractor._KIND_MAP (11 entries — anti_pattern absent), so any such edge vanishes at extraction. Exempt. A negative-only vector is the natural form for an anti-pattern and that is exactly why it must not be authored — it is derivable as the sign-flip of the artifact that rejects it. Authoring it duplicates that vector with an independent drift surface and re-creates the field-authored-relationship shape the repo excised.

[MEDIUM] scripts/generate_schemas.py:4-16 — the design treats "additionalProperties: false on all seven schemas" as seven obstacles. The schemas are generated from the Pydantic models ("the single source of truth"); the flag is a consequence of extra="forbid". One action, not seven. But no generate_schemas reference was found in .github/workflows/ — only tests/doctrine/agent_profiles/test_schema_generation.py:9. If --check is not CI-wired, a field can exist in the model and not in the schema.

[MEDIUM] src/doctrine/directives/models.py:52-68 — Directive has no notes field. notes exists on only 2 of 12 kinds — tactic and procedure. So the existing "Adapted from patterns.sddevelopment.be" credit pattern is structurally unavailable on the majority of kinds that would carry a value vector. Accreditation must live on the value-set artifact, referenced by id — not replicated as prose per artifact.

[MEDIUM] src/doctrine/import_candidates/models.py — CurationImportCandidate and all four sub-models are frozen=True without extra="forbid", unlike LegacyImportCandidate. Unknown keys on the record type meant to carry provenance are silently ignored. Also zero authored instances exist on disk — the type is itself a slot-without-a-producer. The 36-vector import is its first real producer; the design should claim this benefit and does not.

[MEDIUM] creed-and-values-design-as-proposed.md:54-60 — Part 2 straddles two incompatible readings: (a) add impacts alongside in_tension_with, or (b) replace it. (a) is additive and cheap — no new Relation member, so RELATION_DESCRIPTIONS completeness and test_relation_doc_parity do not fire, and DRGEdge.reason already establishes free-text-on-edge precedent. (b) trips both, plus scan_unreconciled_tensions() (src/charter/consistency_check.py:977) and requires an ADR superseding 2026-07-21-1. Pick (a) explicitly and say so in the doc.

[MEDIUM] Pre-existing kind-universe divergence — ten lists: ArtifactKind = 12; _NON_AUGMENTATION_ELIGIBLE_KINDS excludes 3; CHARTER_KIND_TOKENS = 9 + mission-type; NodeKind superset; extractor._KIND_MAP = 11; generate_schemas._REFERENCE_KINDS = 7; directive.schema.yaml reference enum = 9; CharterYaml.activated_* = 8; pack_context valid-kinds = 11 with 10 read helpers; charter/packs/default.yaml:2 claims "all 9 activation kinds" while :5-13 lists 8. This design must not become the eleventh list.

[MEDIUM] §5.6 — if deltas are imported, a hard mandatory-negative constraint cannot be a schema rule at all: it would make 39% of the source corpus unimportable and would invite the importer to invent a negative to satisfy the gate. A fabricated negative is worse than a missing one — unfalsifiable data with the appearance of rigour. Import ⇒ the rule is advisory, never schema minItems.

[MEDIUM] tests/doctrine/test_artifact_compliance.py — an optional field passes; a required one red-lines ~310 artifacts at once. But optional is exactly the shape the 3-for-3 argument condemns. Take optional + a coverage ratchet in the same commit — a test asserting an explicitly-enumerated allow-list of artifacts carries the field, list-only-grows. That converts "partial population" from a silent state into a ratchet.

[LOW] src/doctrine/glossary_packs/ — no JSON schema file exists. A field added there would be Pydantic-validated only — an asymmetric gate. Independent reason to exempt.

[LOW] tests/architectural/test_no_legacy_terminology.py:30-36 — _FORBIDDEN_TERMS covers exactly two retired nouns; _SCAN_ROOTS = ("src","tests","docs") excludes .kittify/. None of the new vocabulary is banned; creed.yaml would be unscanned. Do not represent this change as terminology-gate-verified.

Exemption table — all 12 `ArtifactKind` members

#	Kind	Verdict	Field	Reason
1	`directive`	IN	`value_impact`	A constraint you adopt. "Adopting this rule changes Minimal by −0.25" is exactly the corpus `delta` semantics. Caveat: no `notes` field, so the vector must be self-contained.
2	`tactic`	IN	`value_impact`	Highest-fidelity mapping — the calibration corpus is tactics (36 of 38 practice-shaped). Has `notes`.
3	`styleguide`	IN, caveated	`value_impact`	A convention set you adopt. Caveat: language-scoped and multi-principle, so one vector over five principles is an average. The schema must state the vector is artifact-level, not per-principle.
4	`toolguide`	OUT	—	Descriptive, not adopted. Fields are `tool`/`guide_path`/`summary`/`commands`. You consult a toolguide; there is no Δ-from-not-adopting. Zero of 36 corpus entries is tool documentation.
5	`paradigm`	IN, but	`value_bias`	The design's hidden trap. A paradigm is a value commitment, so its vector is a stance, not a Δ-from-baseline — semantically identical to a profile's bias. The real split is stance-bearing vs adoptable-rule kinds, not "profiles vs everything else."
6	`procedure`	IN	`value_impact`	A gated workflow; running it has costs and benefits. Caveat: deltas dominated by gate cost, so procedures cluster.
7	`agent_profile`	IN, different field	`value_bias`	A profile does not change the system's values by existing; it declares which it weights when acting. Bias, not delta. Hazard: `profile.py:254` lacks `extra="forbid"`, so getting this wrong fails silently.
8	`mission_step_contract`	OUT	—	Structural I/O contract — wiring, not a behavioural rule. A vector double-counts: the same delta counted once on the step and again on the directives its action resolves.
9	`template`	OUT	—	No behaviour of its own. The corpus corroborates: `TEMPLATE_PRACTICE` is one of the 14 all-non-negative vectors — the source library scored a template as pure upside, the tell that templates carry no adoption cost. §5.6 is evidence for this exemption.
10	`asset`	OUT	—	A blob behind a minimal sidecar manifest; content never parsed. Loose contract by design.
11	`glossary_pack`	OUT	—	Terminology, not behaviour — no adoption delta. Sharper: it already carries `provenance: str` as a required field — the only kind that does. If a value vector needs provenance, this is the model to copy, not the kind to extend. Also has no JSON schema at all.
12	`anti_pattern`	OUT as authored	—	Three blockers: not an artifact file (nodes hold 4 fields, no rationale slot); `DRGNode` has no `model_config`; `_KIND_MAP` lacks it. Negative-only is its natural form — hence derivable, never authored.

Net: 6 IN (4 value_impact, 2 value_bias), 6 OUT. Exemption set = _NON_AUGMENTATION_ELIGIBLE_KINDS ∪ {TOOLGUIDE, MISSION_STEP_CONTRACT, GLOSSARY_PACK}, declared once in artifact_kinds.py. The design diverges from the canonical set by exactly one member, and in the direction that is unimplementable.

`value_impact` vs `value_bias` — three arguments

Different referent. A directive's delta is a causal claim about the system; a profile's is a dispositional claim about an actor. Δ(system state) vs a coefficient in a utility function.
Different arithmetic — the killer. Deltas are summable across the active artifact set; that is why delta is the right word. Biases are not: two profiles each weighting Minimal at 0.8 does not give 1.6. A single field name invites the illegal sum and nothing in the schema would catch it.
Different position in the algebra. alignment = creed_weights · artifact_deltas puts a profile on the left and a directive on the right. One shared name guarantees someone eventually dots a profile with a directive.

Corollary the design misses: the creed's field should also be value_bias — the creed is the project's profile. Use different Pydantic types (ValueImpact vs ValueWeighting), not one model with a mode discriminator, so the type system prevents the cross-product structurally (DIRECTIVE_043).

Two useful precedents: AgentProfile.routing_priority (ge=0, le=100) is a scalar bias register on profiles that did not go inert — 18 authored producers plus live consumers. It is the shape value_bias should copy. And model_task_routing's task_fit is the closest existing multi-axis bias vector with a consumer.

Accreditation placement

GATED = must carry a test in the same commit.

The value-set artifact — the single authority. GATED. A loadable artifact with a required provenance field. Copy GlossaryPack.provenance (required) and SourceEntry (name/url/access_method/snapshot_at/license_notes). Fields: value_set_id, name, attribution (required, non-empty), trademark_notice (required), source{url, accessed_on}, authorization_basis. Gate: a validator rejecting empty/whitespace attribution.
Schema $comment — the design's placement is wrong. Schemas are generated; a hand-added $comment is erased by the next regeneration. Author it as Field(json_schema_extra=...), then regenerate.
NOTICE at repo root — new file. GATED. MIT requires retaining the copyright notice; it does not carry third-party trademark notices. Verify it is in the packaged sdist/wheel.
The AMMERSE tactic — close the gap in the same change. notes: with attribution + source URL; references gets the practice-article URL that :61 only gestures at.
Unify the drifted second copy BEFORE accrediting. Two prose copies exist, already divergent ("impact on nature" lost from the template's Environmental definition). Accrediting a misquoted definition is worse than not accrediting.
Generated / rendered output. GATED. If vectors reach agent context, accreditation must ride with the value-set name. The always-on surface is src/specify_cli/tool_surface/profiles/_render_helpers.py:47-65 (13 harnesses, outside the token budget). Cheapest compliant form: "AMMERSE (J.B. Crossland)" once per block.
Docs. One human-readable page, linked from the doctrine index and src/doctrine/README.md.
creed.yaml carries value_set: <id> — a pointer, not inline axis names.

Not sufficient: prose in notes; a hand-edited $comment in a generated schema; a line appended to LICENSE; a mention only under docs/plans/.

Provenance shape

Import, not hand-authoring — and the machinery already exists. import_candidates defines CurationImportCandidate with source{title,type,publisher,url,accessed_on}, classification{target_concepts,rationale}, adaptation{summary,notes}, source_references[], external_references[]{url, attribution_reason, extraction_action}, status, resulting_artifacts[]. external_references[].attribution_reason is literally the field for the AMMERSE trademark obligation.

One import-candidate artifact per import batch, status: adopted, plus on each artifact's value field a two-field origin marker — never a copy of the prose:

value_impact.source_ref: AMMERSE-DELTA-IMPORT-001
value_impact.source_digest: <sha256 of the imported 7-tuple as authored>

Drift detection — three layers:

Content drift (library ↔ doctrine copy). GATED. Promote the corpus JSON to a packaged fixture, then recompute source_digest per artifact and assert a match. A hand-edit of an imported delta then fails immediately. That is the check that would have caught the "impact on nature" loss.
Definition drift (copy ↔ copy in-repo). GATED as a single-authority assertion — assert the seven definition strings appear in exactly one file. A single-authority assertion is refactor-stable; a two-copy diff test rots.
Upstream drift. NOT gateable — accept explicitly. Record accessed_on; never wall-clock-expire (the routing catalog documents why); surface staleness in spec-kitty doctor doctrine.

One constraint to absorb: graph-level provenance does not survive regeneration, so provenance for the impacts edge field cannot live on the edge — it must live on the artifact or in the import record.

Composition with #2216

#2216 is P2 and blocked_by #2467; its all-artefact child is #2591 (component-type on all doctrine artefacts). Sequence, do not batch — and #2591 goes first:

component-type is a discriminator the value field's applicability should be expressed in. Landing the value field first hardcodes a kind-keyed approximation into 7 schemas.
Asymmetric reversibility. component-type defaults to open, absent = today's behaviour. The value field requires 7 numbers + 7 rationales per artifact across ~310 nodes.
Batching shares a failure mode — one schema regeneration, one compliance update, one bulk YAML edit tripping the bulk-edit occurrence gate, one unreviewable diff.
The value field is gated on #2538; component-type is not.

Ordering: #2467 → #2591 → #2538 arm-B → if positive the value field. Hard prerequisite before either: the three silent kind-drop sites. And correct the count: three all-surface sweeps, not two — the impacts edge field is an all-edge sweep the design does not name.

Gates tripped

Gate	Fires?	Detail
`additionalProperties: false` × 7	Yes, one action	Schemas are generated from the models. Confirm `--check` is CI-wired.
`extra="forbid"` on models	Yes — the gaps are worse than the gate	Missing on `AgentProfile`, `DRGNode`/`DRGEdge` (no `model_config` at all), and all five curation import models. A `value_bias` on a profile and an `impacts` on an edge silently vanish today.
Extractor writers	Yes	Field-by-field; fixing models is insufficient.
`CharterYaml` `extra="forbid"`	Yes — blocks `creed` outright	Needs a typed field, not `overrides`. A separate `creed.yaml` needs a place in `hasher.py` + `versioning.py`.
DRG freshness / golden counts	Yes	`tests/doctrine` absent from the ceilings map → implicit ceiling 0. New `== N` needs `# golden-count`.
`RELATION_DESCRIPTIONS` + doc parity	No, if `impacts` stays additive	Fires only if `IN_TENSION_WITH` is retired/reinterpreted — which also needs a superseding ADR.
Glossary integrity	Partly — the miss is the finding	`delta` (= git diff in repo prose), `impact` (vs "AMMERSE Impact Analysis"), `bias`, `value` are ambiguous against existing vocabulary. Per the `primary`/`merge` footgun precedent these need canonical `docs/context/` entries before the tokens reach agent-facing render. Nothing forces a new noun into the glossary — discipline, not a gate.
Terminology guard	No — do not claim verification
Docs freshness	Yes	inventory → index → `--ci`; relative-link check.
`test_artifact_compliance`	Yes	Optional passes; required red-lines ~310. Take optional + a coverage ratchet.
Bulk-edit gate	Yes	~300 YAML files needs an `occurrence_map.yaml`.

Concession

Templates and assets are correctly exempt, and §5.6 is evidence for that. I hunted the counter-example; TEMPLATE_PRACTICE turns out to be one of the 14 all-non-negative vectors — the library scored a template as pure upside, the tell that a template has no adoption cost.
§5.1's insistence that delta is the right word is load-bearing, not pedantry, and it is what let me answer the field-split question cleanly.
Provenance is the design's strongest under-claimed asset. import_candidates has had zero authored instances since WP03; a 36-vector import is its first genuine producer.
routing_priority refutes "3-for-3" as a law. 18 authored producers and live consumers from inception. The rule is a strong empirical regularity, not a structural impossibility, and the discriminating variable is a consumer in the same seam. I will not carry 3-for-3 as a blocker against this design.
Part 2 is cheaper than feared, provided it stays additive.

Verdict

The design keeps the artifact catalog coherent only if three things change, each a small edit rather than a rebuild. First, the exemption boundary must be code, not prose — it diverges from the canonical set by omitting anti_pattern, the one kind structurally incapable of carrying the mandatory rationale, and five more are ruled out on semantics, leaving 6 in and 6 out; as a prose sentence this becomes the eleventh divergent kind enumeration in a codebase whose canonical kind file forbids a second one. Second, agent_profile needs a genuinely different field — value_bias, a distinct Pydantic type — because deltas compose by summation and biases do not, and the operation this design enables puts profiles and directives on opposite sides of a dot product; the design also does not anticipate that paradigm belongs on the profile side of that split. Third, accreditation must be a required field on a loadable value-set artifact, because prose is unavailable where needed (notes on 2 of 12 kinds), LICENSE says "Copyright GitHub, Inc.", there is no NOTICE, and a hand-added schema $comment is erased by the next generation — and the unification of the two drifted definition copies must land first, or you accredit the drift. On sequencing: three all-surface sweeps, and #2591 goes first. Finally, the two gates most likely to let this ship inert are not the ones the design names: AgentProfile lacks extra="forbid" and DRGNode/DRGEdge have no model_config at all, so a value_bias on a profile and an impacts on an edge both load and silently vanish today — and the extractor writers would drop them again at regeneration even after the models are fixed.