Datasets:
| # Verifier entrypoint — Harbor runs this after the agent finishes. | |
| # | |
| # Grades the delivered redline (/app/contract.docx, edited in place by the agent) against this task's | |
| # weighted rubrics (tests/rubrics.json) using: | |
| # 1. the validity gate (forgery / integrity / did-real-work checks), | |
| # 2. the change-list parse (baseline vs delivered → a trusted, forgery-resistant edit record), | |
| # 3. a 3-provider LLM judge panel (majority >=2 of 3), then the signed-weight score. | |
| # | |
| # The pristine turn input (grader-only baseline) is baked at /tests/baseline/contract.docx. | |
| # Panel composition + API keys come from env (carried by task.toml [verifier.env]): | |
| # JUDGE_PANEL="openai/...,anthropic/...,gemini/..." (default panel used if unset) | |
| # OPENAI_API_KEY / ANTHROPIC_API_KEY / GEMINI_API_KEY | |
| set -uo pipefail | |
| python3 /tests/judge.py \ | |
| --contract /app/contract.docx \ | |
| --baseline /tests/baseline/contract.docx \ | |
| --out-dir /logs/verifier | |