CrewRift Player Optimizer Learning Package Review

Branch review for extraction from co-gas sessions, files, candidate ledgers, and tooling.

Review Scope

Generated package Metta branch review

Output

The bundle is staged in this branch at:

packages/coworld/templates/optimizers/242026/xpser-co-gas-crewrift-player-optimizer

Review branch: relh/xpser-crewrift-learning-package. Review machine: xpser.

Workflow from PR branch

Workflow Source

The requested skill was not present on local Metta main after fetch. I read it from PR 15623 head:

origin/pr-15623-extract-learning-package

Requested file:

docs/ai/onboarding/workflows/extract-learning-package/SKILL.md

Prepared for Phase 5

Likely Destination

This package is already in the current-week optimizer package area for branch review:

packages/coworld/templates/optimizers/242026/xpser-co-gas-crewrift-player-optimizer

This review page is the sign-off boundary before opening PR work.

Phase 1: Session Mining

Session mining performed Partial parse warnings

Mining Stats

  • 3,782 session or history files were enumerated by the workflow helper.
  • 290 files matched broad CrewRift, co-gas, Coworld, XP, tournament, or Observatory terms.
  • 22 current Codex sessions in this repo were narrowed and read for user/final-turn signal.
  • One older Claude exploration under /home/relh/gt/co_gas contributed historical context.
  • Some JSONL session files were invalid or truncated, producing parse warnings. Those files were not silently treated as complete.
4 session-derived learnings

Recovered Session Lessons

  • The optimization loop matured from local scrimmage-based promotion into remote Observatory XP evidence gates.
  • Slot-conditioned CrewRift strategy was corrected: slots are runner and protocol metadata, not role, identity, or strategy.
  • Source custody became a release requirement: tournament policies must be buildable from repo-local source.
  • Local runs are diagnosis and smoke checks; completed hosted XP artifacts are the replacement evidence.
No raw transcripts included

Transcript Handling

The package preserves distilled lessons, not raw transcripts. Session-derived material is labeled as reconstructed and should be reviewed against the source-backed files before promotion into shared docs.

Draft Loop Summary

The reconstructed loop is: refresh mandate, select the lower owned lane and highest-impact failing game, inspect concrete bad episodes, name a behavior failure, edit source-backed policy code, run short local smoke checks, mark the candidate local_passed, create a no-submit hosted version for the lower lane, request XP against leaders, lower champion, and bad matchups, inspect replay/log/artifact evidence, then promote only when completed evidence supports replacing the lower champion.

Phase 2: Swept Files And Haul

Haul inventory created

Haul Files

  • haul/file-agents.md: active mandate, CrewRift slot rule, submission gates.
  • haul/file-readme.md: canonical commands, lanes, source entrypoints.
  • haul/file-playbook.md: workflow doctrine and durable tournament practice.
  • haul/file-crewrift-suspectra.md: CrewRift social memory and Suspectra behavior model.
  • haul/file-crewrift-candidates.md: candidate evidence from v98, v99-v102, v108, and split-hold records.
  • haul/file-tooling.md: co-gas CLI, XP request helpers, source upload, frontier analysis, and local diagnostic scripts.
Evidence status matters

Performance Record

performance/LOG.md records that crewrift-suspectra-richard:v98 is the strongest completed-XP-supported replacement found in the swept evidence. Later v99-v102 pressure variants were rejected by XP evidence, and v108 remained a no-submit hosted candidate without completed XP replacement evidence.

The latest mandate snapshot still shows CrewRift unmet: public leader RowDaBoat rank 1, RelhAlpha rank 13, Richard rank 14, with only Richard tracked as champion.

Null tier created

Cut Material

null-tier.md excludes raw transcripts, stale local-only promotion habits, slot-conditioned strategy, pending or running XP rows, stale wrappers, and full historical ledgers that would bloat a teaching package.

Phase 3: Universal Package

Universal tier

Scope

The universal tier keeps principles that should transfer to many tournament optimizers: evidence-gated iteration, failure-led edits, artifact inspection, source custody, local smoke limits, and explicit hold decisions.

Files: universal/AGENTS.md, universal/LOOP.md, universal/MANIFEST.md, universal/performance/LOG.md.

Skill included

Reusable Skill

universal/skills/evidence-gated-iteration/SKILL.md teaches the general loop without co-gas-specific commands. It emphasizes connecting one observed failure to one correction and one measured outcome.

Guides included

Guides

The universal guide covers evidence-led optimization. A second guide preserves transferable social-game lessons while avoiding CrewRift-only mechanics claims.

Phase 4: Tiered Packages

Fully general

Fully General

fully-general/ contains lessons that should apply outside Coworlds: causal diagnosis, benchmark hygiene, source custody, artifact inspection, local versus remote evidence boundaries, and null-tier pruning.

CrewRift-specific

CrewRift-Specific

crewrift-specific/ focuses on role-aware social play, suspicion and vote analysis, alibi quality, kill conversion, task/fake-task behavior, and the canonical slot rule. It states that mechanics claims must be checked against cogas-agents/coworlds/crewrift/vendor/coworld-crewrift/src/crewrift/sim.nim.

Loop-specific

Co-Gas Remote XP Loop

co-gas-remote-xp-loop-specific/ keeps exact command patterns, lane discipline, candidate ledger practice, source upload expectations, and Tournament Observatory XP request workflow.

Phase 5: Sign-Off Boundary

Human sign-off required

Not Yet Uploaded

The workflow's PR phase is intentionally not complete. Review this page and the package files first. After sign-off, the next mechanical step is to open the PR requested by the workflow from relh/xpser-crewrift-learning-package.