Critique and Revision

Constitutional-AI-style supervised phase: draft -> critique against a written principle -> revise. The constitution here is a single quality principle ("state uncertainty honestly; never present a guess as a fact"). A decisive drafter answers, a temperature-0 critic judges the draft only against the principle, and violating drafts are rewritten. Every row carries full provenance (principle, critique, verdict), so downstream curation can trace exactly why each response looks the way it does.

Files

• basic.py - draft -> critique -> revise over 5 prompts: Fermi estimates and predictions that tempt overconfident guessing, plus settled facts that should pass untouched. Writes SFT rows where the response is the revision when the critic found a violation and the original draft otherwise.
• constitution_pairs.py - same loop over 6 prompts, but emits preference pairs (chosen=revision, rejected=draft) only where the critic found a violation and the revision actually differs from the draft. Ends by printing a paste-ready list literal in the exact EXAMPLES shape of ../_05_text_pairwise_preference/dpo_jury.py: this folder generates the pairs that the _05 jury consumes, so self-generated pairs get relabeled by a jury. The pairs carry source_family="google" (the drafting model's family), so dpo_jury's self-preference recusal benches its Gemini juror on them and only the other four families vote. Chosen/rejected alternate between slots a and b by index parity, so a jury with position bias is exposed.
• iterative_revisions.py - adds a temperature-0 judge (score 1-5 against the principle plus general quality) as a stopping rule: revise only while the score is below 4, up to 3 judge rounds (so at most 2 revisions). Rows store the score trajectory per prompt - the calibration signal for how many revision rounds a production pipeline should pay for.

Rows are written to data/generated/ (gitignored - run the scripts to regenerate). Abridged rows from a real run:

{"prompt": "How many piano tuners work in Chicago?", "response": "There are an estimated 70 to 100 professional piano tuners currently working in Chicago. While the exact number fluctuates and is difficult to verify precisely, ...", "provenance": {"principle": "State uncertainty honestly. ...", "violates": true, "critique": "The draft presents the exact number of 82 piano tuners as a certain fact, failing to acknowledge that this is an estimated or variable quantity ...", "revised": true}}
{"prompt": "How many stars are in the Milky Way?", "chosen": "The Milky Way galaxy is estimated to contain between 100 billion and 400 billion stars, as the exact number cannot be directly counted.", "rejected": "The Milky Way galaxy contains 100 billion stars.", "provenance": {"principle": "State uncertainty honestly. ...", "critique": "The draft presents the estimated number of stars in the Milky Way as a definitive fact (\"100 billion stars\") instead of acknowledging it as an estimate ..."}}
{"prompt": "When will fusion power be commercially widespread?", "final": "It is highly uncertain when fusion power will become commercially widespread, with estimates ranging from 2050 to late in the 21st century. ...", "rounds": 2, "score_trajectory": [1, 5]}

When to use

When you have a written quality or style principle and want the model to enforce it on its own outputs:

• SFT distillation: basic.py rows train the principle into a model without the principle in the prompt at inference time
• Preference data: constitution_pairs.py pairs feed DPO or reward model training, with the critique as the documented reason
• Quality gating: iterative_revisions.py when one revision pass is not reliably enough and you need a measured stopping rule

The pairs this folder generates are consumed by the jury in _05_text_pairwise_preference/ - run constitution_pairs.py, paste the printed list into dpo_jury.py's EXAMPLES, and let independent judges relabel them. If you can verify outputs mechanically instead of critiquing them, use _21_rejection_sampling/. To filter or dedupe the resulting corpus at scale, use _22_dataset_curation/.

Run

python cookbook/data_labeling/_23_critique_and_revision/basic.py
python cookbook/data_labeling/_23_critique_and_revision/constitution_pairs.py
python cookbook/data_labeling/_23_critique_and_revision/iterative_revisions.py

Requires GOOGLE_API_KEY.