Learning From Votes¶

The feedback loop lives in src/casita/llm.py and the CLI in src/casita/__init__.py.

Votes and pass reasons are stored in SQLite. During ranking, Casita builds:

• inline feedback for listings in the current batch
• a capped few-shot block of recent up/pass examples
• an audit prompt exposed through casita analyze-prefs

analyze-prefs reads the votes and compares revealed preference against the static ranking policy. It proposes contradictions and new rules, but it never edits code. A human decides whether a proposed rule belongs in the prompt.

Ways This Could Go Further¶

The loop could gain better fixtures, better diff output, or clearer aging of old examples. The important property to preserve is reviewability: revealed preference should become policy through an intentional code change.