Human Adjudication Ingestion

scripts/human_reliability/ingest_adjudication.py validates one JSON adjudication submission against the #86 contract and the exact frozen #85 queue:

npm run human-reliability:adjudicate -- \
  --case lincoln \
  --cohort lincoln-en-cmt \
  --cohort-version 1.0.0 \
  --json /secure/adjudication-submission.json

Immutable registration

Input bytes receive a deterministic adjudication-registration-* ID and are stored read-only beneath adjudication/decisions/raw/<registration_id>/adjudication.json. Re-ingesting identical bytes is idempotent and verifies the stored bytes. Every attempt, including invalid attempts, remains in adjudication-register.json with a JSON/Markdown validation report.

Only one valid submission may govern a frozen queue digest. A corrected submission may follow an invalid attempt, but a second valid decision set for the same queue is rejected rather than silently superseding the first.

Queue and decision reconciliation

Validation checks:

  • case, cohort, language, layer, queue source, digest, schema, generator, generator hash, and code revision;
  • every stable queue ID exactly once, with matching disagreement, item, reference, unit, and field identities;
  • unique adjudication, queue, evidence, claim, and correction-candidate IDs;
  • adjudicator membership declarations against the approved cohort;
  • status-specific values, selected coder/reference basis, controlled vocabularies, bounded confidence, and nontrivial rationales;
  • exact affected-claim and affected-dimension coverage from the queue; and
  • correction targets against the queue reference ID, field, current reference value, adjudicated value, and case namespace.

Invalid submissions are preserved and reported but never enter normalized results.

Normalized results and unresolved cases

Valid decisions are joined to their complete queue entries in adjudication/results/<cohort>-<version>/adjudication-results.json. Original coder values, reference summaries, model summaries, source context, priority, and the adjudicator’s decision remain separately recoverable.

deferred and unresolved decisions remain first-class normalized results. They set the cohort adjudication state to unresolved; ingestion never fills a missing value from a coder majority, reference match, or model summary.

Dedicated correction review layer

Only schema-valid candidate records are copied into correction-candidates/<cohort>-<version>/correction-candidates.json and CSV. The layer is explicitly review-only:

  • promotion_permitted is false;
  • promotion_status remains pending_separate_authorization;
  • promotion_id remains null; and
  • direct_write_permitted remains false.

The command writes only below quality/human-reliability/. It does not open accepted annotations, corpus files, analysis artifacts, metadata, claims, or publication files for writing. Correction promotion remains a separate, authorized workflow.

Adjudication evidence may reproduce source-derived material and inherits the approved cohort’s repository_allowed or local_only policy. Local-only case subtrees must remain gitignored.