Adopt ADMIT-009: extraction_confidence must not persist onto admitted canon

notations/CONTRACT.md

@@ -201,6 +201,7 @@ For `admission_state: proposed`, the §6 requirement on `admitted_at` / `admitte
 | `ADMIT-006` | error | `reviewer_authority` is present and not one of `ai_reviewed` / `expert_confirmed` (§6.2). |
 | `ADMIT-007` | error | The admitter does not match the tier: `reviewer_authority: ai_reviewed` but `admitted_by` identifies a human, or `reviewer_authority: expert_confirmed` but `admitted_by` identifies a tool. A tool writes `ai_reviewed`; a human writes `expert_confirmed` (§6.2). |
 | `ADMIT-008` | warning | `admission_state: proposed` and `owner_to_confirm` is absent — the open item has no designated reviewer and will not route to any inbox. |
+| `ADMIT-009` | warning | An admitted artefact (`admission_state: active` or absent) in the `canon` zone carries `extraction_confidence` — a review flag that belongs to an ingest candidate and is never persisted into canon (§11.1). Cite the provenance through `derived_from` and the field artefact it came from instead. Does not apply to `canon/unresolved/` (§13, untyped) or to candidates under `_intake/`, where the flag is correct. |
 
 This lifecycle is what an automated regulatory-intelligence collector (a separate task) depends on: the collector emits `proposed` drafts plus a review digest, and the human gate admits or rejects. The per-TYPE specs note it where relevant ([15-requirement.md](elements/15-requirement.md), [16-assertion.md](elements/16-assertion.md)).
 
@@ -646,6 +647,8 @@ Confidence is **metadata about certainty, not a contradiction flag.** A low-conf
 
 The two are deliberately separate. Collapsing them into one decaying number would erase the authored signal. A statement's overall confidence combines them (§11.4).
 
+**A third signal exists upstream and never reaches canon — `extraction_confidence`.** It answers a different question from either of the two above: *did the extractor read the source document correctly*. It is a review flag on an ingest **candidate** (`candidate.extraction_confidence`, [vocabulary.yaml](vocabulary.yaml)), surfaced in the review queue and used to route reviewer authority (§6.2). It is **never persisted into canon**, is never folded into `source_quality`, and never enters the §11.4 formula: once a candidate is admitted, the candidate is gone, and how confidently a now-superseded extraction was read is not a property of the element. An admitted element carrying it misrepresents a one-time routing signal as a permanent, citable attribute — the provenance an admitted element does carry is `derived_from` plus the cited field artefact (an `OBSERVATION` or other source record, [IDS_AND_REFERENCES.md](IDS_AND_REFERENCES.md) §3.4). Enforced by `ADMIT-009` (§6.1); stated per TYPE in [ELEMENT_PRIMITIVES.md](ELEMENT_PRIMITIVES.md) §7.29.
+
 ### 11.2 Source-trust scale
 
 `source_quality` is a closed ordinal set. The numeric weight is used only for the arithmetic in §11.4–11.6; authors record the label, never the number.

notations/ELEMENT_PRIMITIVES.md

@@ -1015,7 +1015,7 @@ A `RELEASE` is a dated/versioned **state of a modelled subject** — one shipped
 
 **Provenance — no per-TYPE fields, because the envelope already carries it.** A `RELEASE` admitted from tag/branch evidence rather than a signed release document cites that evidence through **`derived_from`** (§3) — an envelope field every `standalone` TYPE carries, not one a TYPE opts into. There is consequently nothing for this section to add and nothing missing: a tag-sourced release is exactly as auditable as a document-sourced one, and its provenance belongs there rather than buried in free-text `description`. The matching field artefact for a repository/tag survey is an `OBSERVATION` ([IDS_AND_REFERENCES.md](IDS_AND_REFERENCES.md) §3.4).
 
-`RELEASE` has **no** `extraction_confidence` field — and neither does any other canon element, of any TYPE. `extraction_confidence` is a review flag on an ingest **candidate** (`candidate.extraction_confidence`, [vocabulary.yaml](vocabulary.yaml)); it is surfaced in the review queue, drives reviewer-authority routing ([CONTRACT.md](CONTRACT.md) §6.2), and is **never persisted into canon**. An admitted element carrying it is a defect in whatever wrote it, not a precedent to follow. (Decision recorded per `transitrix-hq#197`: documented convention, no schema extension.)
+`RELEASE` has **no** `extraction_confidence` field — and neither does any other canon element, of any TYPE. `extraction_confidence` is a review flag on an ingest **candidate** (`candidate.extraction_confidence`, [vocabulary.yaml](vocabulary.yaml)); it is surfaced in the review queue, drives reviewer-authority routing ([CONTRACT.md](CONTRACT.md) §6.2), and is **never persisted into canon**. An admitted element carrying it is a defect in whatever wrote it, not a precedent to follow — flagged by `ADMIT-009` ([CONTRACT.md](CONTRACT.md) §6.1), which points the author at `derived_from` and the cited field artefact instead. (Decision recorded per `transitrix-hq#197`: documented convention, no schema extension.)
 
 ```yaml
 # canon/elements/05_implementation/releases/RELEASE-PAYMENTS-GATEWAY-2.yaml

notations/vocabulary.yaml

@@ -598,6 +598,9 @@ rule_codes:
   ADMIT-008:
     severity: warning
     spec: notations/CONTRACT.md
+  ADMIT-009:
+    severity: warning
+    spec: notations/CONTRACT.md
   AGREE-001:
     severity: error
     spec: notations/CONTRACT.md