Public handling: selected full-text prompt with secret filtering.
Prompt Text
You are auditing DoneBench task quality for one domain. Workspace: D:\Research\DoneBench. You are not alone in the codebase: other workers are auditing other domains and writing disjoint files, so do not revert or modify their work.Context: A GPT-5.2 full-domain audit was attempted but the model route returned 503; this run uses the current available strong model path and must NOT be described as GPT-5.2. This is model-assisted audit, not human annotation.Your ownership: email domain only. Read annotation/human_audit_queue.jsonl, annotation/annotation_guide.md, reports/audit_full_domain_model_assisted/README.md, prior relevant audit evidence in reports/audit_deepseek_gpt55_merged/, and data/tasks/email/email_021.json through email_040.json.Write exactly one output file: reports/audit_full_domain_model_assisted/email_ai_audit_opinions.jsonl. It must contain 20 JSONL records, one per email_021..email_040 task, using the schema in the README. Use prompt_version "full-domain-model-assisted-v1", model "model_assisted_full_domain_audit", provider "model_assisted_audit", provider_model "current-available-strong-model". Do not edit annotation/human_audit_queue.jsonl or human fields.Audit checks: criteria_complete, donespec_matches_criteria, near_misses_are_valid, reference_trace_is_plausible, not_too_templated. Verdicts pass/warn/fail with concise rationales and confidence. overall_risk low/medium/high. Set needs_adjudication true for high risk or material unresolved semantic gaps. Be rigorous; do not try to make the gate pass. Independently verify prior high-risk claims from task JSON.When done, final-answer with the path written, count of records, high-risk task_ids, adjudication task_ids, and notable domain pattern.
Reuse Notes
EXTRACTED: This is a selected Codex prompt or automation prompt from the local Codex corpus.
INFERRED: Future agents can reuse its structure, constraints, and acceptance criteria when creating similar Codex workflows.