Human Adjudication Queue Generation

scripts/human_reliability/generate_adjudication_queue.py converts a validated human disagreement log into a deterministic queue for an independent, authorized human adjudicator:

npm run human-reliability:queue -- \
  --case lincoln \
  --cohort lincoln-en-cmt \
  --cohort-version 1.0.0

The command writes adjudication-queue.json and adjudication-queue.csv beside the cohort’s disagreement log. The JSON output conforms to schemas/human-reliability/adjudication-queue-schema.json.

Evidence boundary

Each entry preserves the disagreement and source-pattern IDs, both coder values, the separately labeled reference summary, category, claim impact, source-language risk, priority reasons, and bounded review question. Source text and focal lexical text come only from the original blind packet after its manifest self-hash and payload hash are verified.

Direct affected claims are linked through matching unit, sentence, or reference IDs in an available claim-trace artifact. Broader affected-claim dimensions such as metaphor presence, mapping, sacrifice, purification, violence, obligation, agency/absence, and source-language interpretation remain separate from direct claim links.

A valid model-reliability agreement artifact is optional. When present, only matching aggregate field summaries are included. When absent, the queue records an explicit unavailable status. Model summaries are diagnostic context, not votes, and are never blended with coder values or reference evidence.

Priority policy

The queue begins with #84’s adjudication priority and then applies deterministic increments. It promotes:

  • a presence/nonpresence disagreement;
  • both coders diverging from the reference;
  • high claim impact or a direct claim trace;
  • agency/absence, purification, and violence fields; and
  • material/high source-language risk or an uncertainty split.

Ranks may change when new disagreements or claim traces are added. Queue IDs are hashes of stable disagreement IDs, so downstream adjudication records can refer to the same item across re-ranking.

Safety and authority

The queue contains no adjudication decision, accepted value, or reference update. It is marked pending-independent-adjudication, and decision authority is reserved for an authorized human adjudicator. Generation writes only under the cohort comparison directory and never modifies packets, submissions, accepted annotations, model artifacts, claim traces, or source corpora.

Queue files reproduce packet source text and therefore inherit the cohort’s repository_allowed or local_only storage policy and rights constraints. Local-only queues remain local and untracked with their source-derived cohort artifacts.