Research acceptance and exclusions

Authority and current status: DEC-0022 (repository reference). Accepted on 7 October 2026 through that decision after independent normative promotion review. Acceptance adopts the exact payload below, not every surrounding proposal or research file.

Exact normative payload

The RE01-RE08 source text is docs/specification/extensions-draft/research-estimates.md at 4eb5e35935fb98b5f798429fe181aab3882d7266. Its byte range begins with ## RE01 and ends immediately before The [NBER dossier], including intervening line breaks. Introductions, drafting status and trailing readiness prose are excluded. That same payload remains in its source file for inspection and is protected by the hash below.

One sentence in RE07 is informative only and is excluded from normative adoption:

Use the proposed change contract for revisions, without treating a changed estimate as economic evolution or proof of a changed method.

RP01 separately retains the prohibition on inferring economic or method change from a changed estimate. It does not require the proposed change contract. No CH, IF, FC or DS clause is adopted. RE08's review dimensions remain editorial labels; its disclosure and operation-boundary requirements are adopted within this profile.

The profile boundary consists of RP01-RP04 in README.md, beginning with ## Normative boundary and ending immediately before ## Evidence and deferred work. These requirements explain minimum disclosure, unknowns, full versus restricted review, documentation labels and separate operation scope. Its introduction and research agenda are informative. The profile boundary governs the explicit adoption exclusions above.

Adopt all 31 cases in docs/corpus/research-v0.1/inputs.json together with the paired expectations.json. Each case's inlined source material, candidate, restrictions and scope are its normative premises; its documentation finding, decisive reason, material secondary findings and separate operation disposition are its scoped oracle. Equivalent explanations are acceptable. RC01 covers full minimum disclosure on a partial record; the other cases exercise named provisions only. Their success cannot establish unexercised conformance.

The cases' historical candidate labels and JSON field names are editorial. This decision adopts their logical obligations and case expectations, not a canonical representation or an implementation that copies expected outputs. Old dossiers and corpus artifacts retain their original status. No complete dossier mapping is automatically made authoritative.

Artifact identity

SHA-256 pins byte identity. It is not proof of economic meaning, publisher authenticity or statistical truth. The two source dossiers used for case projections remain pinned inside each projection and unchanged from the base. Whole-file pins apply to the case/oracle JSON; the rule/profile pins apply to the exact UTF-8 ranges defined above.

Required source material and promotion gate

The case protocol (repository reference) requires two separate read-only classifiers to reconstruct documentation findings and operation dispositions without the oracle as input. Freeze the oracle before dispatch. Preserve initial outputs, source/rule hashes and disagreements in the review record (repository reference). Record any corrective change and repeat its affected reviews; do not overwrite a first-pass disagreement.

Before acceptance, obtain independent review of the complete normative promotion candidate, including exact scope, exclusions, cases and actual classification source material. Root records an integration disposition only after inspecting those reviews. Verify the final clean SHA, allowed paths, source/rule/case pins, unchanged core artifacts and green repository checks. Integrate through a private PR, verify the merged tree and checks, then prune only this work.

Fixture tests verify integrity, source projections, input/oracle separation, required groups and recorded review linkage. They do not implement a classifier, certify methodology or measure LLM extraction accuracy. Published-source familiarity and correlated agent errors remain possible. A later implementation must classify isolated case inputs itself.

Classification and promotion record

Initial classification at b024852cb65590702862069d648c80cc34b5e0ba is preserved in the review record. Both reviewers agreed on 30 findings; one identified RC23's missing prose-unit premise. A separate preflight identified RC13/RC14's incomplete locator context and the missing rule-hash guard. Those issues were corrected without changing the normative rule payload or expected findings. Both reviewers repeated RC13, RC14 and RC23 at a506f8c0115b9dc62f48672cb641cbd30674840b, verified the other 28 complete cases unchanged, and retained those prior judgments. All 31 corrected findings and decisive distinctions agree with the oracle. Whole-file review hashes and root-computed normative-range hashes bind the source material to the current payload.

Both research_adoption_preflight and research_claim_classifier_b accepted normative promotion at f6f62e1816707c1011753fab27a9aa08505822a6. Root records the decision through private PR #28 under the user's delegated authority. The final status/index/plan updates receive exact-head review before merge. The rule and case payloads remain unchanged. Full checks at the reviewed promotion head passed 50 repository and 253 SDMX tests; final integrated checks are recorded on the PR and in the workstream run record.

The adopted oracle contains 12 meets, 18 violates and 1 not_assessable documentation findings. RC24-RC26 additionally require SC08 out_of_scope. Those counts describe the bounded oracle, not measured classifier performance.