Judges (SimpleAudit 0.2.3: customize_judge / build_judge)
- A judge version is criteria (what to evaluate) in an output format,
plus the probe prompt: start from a SimpleAudit judge (its format kept)
or use a generic severity, score 1-10 (own dimensions, averaged in code)
or yes/no-question format. Replaces rubric + raw judge prompt.
- Unedited SimpleAudit judges still run by name; edited or own criteria
run as composed judges. Migration converts versions and run snapshots.
- Editor: "Start from" picker, format panel, criteria with reset, probe
prompt, live full-prompt preview; text follows the start when untouched.
- Clone opens a prefilled new-judge form (nothing saved until Create);
SimpleAudit judges with a matching generic format clone as own criteria.
- What changed between any two versions; delete unused versions and
unused judges from the judge page.
Versioning, one vocabulary and shared components
- Versions (scenario sets, judges) are "v4" / "v4 · latest" every…
