gepa-frontier

gepa-frontier

After a run completes (done.json exists), inspect the candidates and — only on deliberate request — write the winner back onto the artifact. Promotion is never silent: show the diff, get the nod, then commit.

Inspect

gepa-anywhere frontier --config .gepa/config.yaml

Prints:

• frontier — every evaluated candidate with its per-example scores and mean, sorted best-first. Look for a candidate that wins broadly (high mean) AND doesn't tank any single example (per-example scores) — a specialist that collapses one example is worse than a slightly-lower generalist.
• done — best_idx, the holdout summary (seed_mean → best_mean, delta), and the cost note.

Read the holdout first. If best_mean did not beat seed_mean on the holdout split (delta ≤ 0), the gain was over-fit to train/val — do NOT promote; tell the user the run didn't generalize and suggest more/cleaner golden data or a sharper metric.

Promote (deliberate)

When the user confirms they want the winner applied:

1. Show the change first. The current artifact is at artifact.path; the winner's text is in done.json → best_candidate. Display a diff so the user sees exactly what changes.

2. On confirmation:

   gepa-anywhere frontier --config .gepa/config.yaml --promote
   

   This writes the winning candidate's text onto the artifact file(s) and prints what it wrote. It does NOT commit — leave that to the user (or, if asked, git add + a reviewed commit).

3. If the user prefers, you can promote by hand (copy best_candidate.<component> onto its file) — same effect, fully reviewed.

Guardrails

• Never promote without showing the diff and the holdout delta.
• Never promote a run whose holdout regressed.
• best_candidate keys map to the config's components (single_file → prompt). Promotion ignores any key that isn't a real component.