Crewrift Player Optimizer Learning Package Review

Source file: haul/LOOP.md

1. Start from AGENTS.md and the player_labs guide.
2. Confirm RelhAlpha and resume on crewborg if already recorded.
3. Run uv sync and the crewborg tests before changing behavior.
4. Build and upload a baseline, then append version details to version_log.md.
5. Query live Crewrift state before choosing evaluation opponents.
6. Create a schema-checked hosted experience request, using enough episodes to observe behavior.
7. Fetch all artifacts with an explicit count.
8. Inspect artifacts, make one focused policy change, rebuild, and upload.
9. Verify the submitted policy belongs to the intended player; use the player login route if the CLI defaults to the wrong identity.
10. Report live standings side by side for RelhAlpha and the Richard Higgins policy lane.
11. Repeat the experience -> inspect -> change -> measure loop.

Do the mechanical work, narrate plainly, and pause only for real human choices, login, or sign-off.

Persist crewborg and RelhAlpha in the lab preferences so later sessions do not restart the starter-choice interview.

Query current league, divisions, leaderboard, and active memberships before creating hosted experience requests.

Verify completed episode count and pass an explicit -n; the downloader default fetched only 10 of 32 episodes.

If the CLI assigns policy versions to the wrong player, inspect OpenAPI and use the player login route without logging tokens.

Report RelhAlpha and Richard Higgins policies side by side using current leaderboard plus active membership data.

Write scratch request bodies, run local schema checks, and catch bad Crewrift config shapes before creating hosted work.

The run started with RelhAlpha:v1, completed a 32/32 hosted baseline request, produced RelhAlpha:v2, and ultimately submitted RelhAlpha:v4 under the correct player identity. The session's final live snapshot had RelhAlpha rank 1 and Richard Higgins rank 4, with the round still moving.