coach-improve-pco

name	coach.improve.pco
description	PCO improvement iteration. Plays a game, collects experience, runs CvCLearner to propose program patches, tests locally, and submits if improved. Use when asked to "run PCO" or "evolve programs".

PCO Improvement

Run one PCO (Program Conditioned Optimization) cycle: play → collect experience → learner proposes patches → test → submit.

Reads eval/submit commands from IMPROVE.md.

Steps

1. Eval Baseline

Run eval from IMPROVE.md on seed 42 to establish baseline score.

2. Play a Game & Collect Experience

rm -f /tmp/coglet_learnings/*.json

Run eval again (experience is written to /tmp/coglet_learnings/ automatically).

3. Run PCO Epoch

import asyncio, json, glob, anthropic
from cvc.pco_runner import run_pco_epoch
from cvc.programs import all_programs

f = glob.glob('/tmp/coglet_learnings/*.json')[0]
experience = json.load(open(f))['snapshots']

result = asyncio.run(run_pco_epoch(
    experience=experience,
    programs=all_programs(),
    client=anthropic.Anthropic(),
    max_retries=2,
))

Log signals (resource, junction, survival magnitudes) and proposed patches.

4. Review & Apply Patch

If result["accepted"] and patch looks reasonable:

• Fix any invalid API calls (the learner sometimes invents methods)
• Apply to programs.py — only modify the specific function
• Valid GameState API: see IMPROVE.md

If not accepted, log why and report back to coach.improve (it may switch to design).

5. Test Across Seeds

Run eval across 5+ seeds. If average score drops vs baseline, revert the patch.

6. Submit if Improved

Use the submit command from IMPROVE.md. Log the submission version.

Output

Report back with:

• accepted: whether PCO produced an accepted patch
• improved: whether scores improved after applying the patch
• score_before: baseline average
• score_after: post-change average (or null if reverted)
• signals: loss signal summary from PCO
• description: what changed