Reproduce the model-judge phase

Wolfram replaced the human-rating phase with blinded OpenRouter judges. There is no participant recruitment or manual coding task to complete. The former human instructions remain an archive.

1. Read the judge protocol, including transport and parser amendments.
2. The 24 captured conversations are fixed. The task manifest contains anonymous paired text, both orders, and deduplicated openings.
3. Run .venv/bin/python probes/folk01_judge.py run with the local OpenRouter key. Completed exact requests are reused; missing requests consume API credit. Never publish the key.
4. Run .venv/bin/python probes/folk01_judge_analyze.py for the separate judge tables, uncertainty bounds, order checks and diagnostic associations.
5. Run .venv/bin/python probes/folk01_judge_audit.py for quote sensitivity and the arithmetic check. Run .venv/bin/python probes/folk01_judge_verify.py to check coverage, schema extraction, request hashes and quote flags.
6. Regenerate with .venv/bin/python probes/folk01_site.py, then the standard site and preview generators.

Read the report or inspect raw model judgments. Model judgments stay separate from the former human importer. No synthetic human responses are created. The remaining two frozen topics are held out for future predictor validation.