Crewrift Player Optimizer Learning Package

The transcript pile was too large for full-context reading. The workflow helper found 29260 sessions. A direct Crewrift search found only this current extraction request. The draft loop therefore uses file-backed Crewrift evidence plus analogous policy-diagnostic sessions from March 2026.

Run episode -> collect artifacts -> expand replay/logs -> name one failed capability -> patch smallest owner -> add narrow regression -> rerun same evidence path -> compare -> upload only after baseline comparison.

No direct Crewrift policy score trajectory was found. The performance log intentionally says this rather than inferring scores from process or git history.

Crewrift optimization should start from replay timelines. The replay expander exposes phases, movement, tasks, kills, bodies, reports, votes, chat, score changes, and optional JSONL snapshots.

Notsus provides the capability map: Sprite v1 protocol, map/walkability, role/color detection, pathing, task holding, body reporting, imposter behavior, voting, and chat.

The optimizer is an interactive workbench. Use local run-episode/play/replay/scrimmage loops before hosted XP, and treat uploaded policy versions plus evaluation evidence as outputs.

Session candidates reinforce targeted scenario tests, parser invariants, trajectory reports, and small owner patches, but they are not direct Crewrift evidence.

Universal package routes always-on heuristics to AGENTS.md, concrete process to LOOP.md, replay and capability recipes to skills, Crewrift/Coworld detail to guides, and source tool pointers to tools/README.md.

Always start from evidence. Name one failed capability before editing. Patch the smallest owner. Keep the loop replay-driven. Mark session-derived lessons until reviewed. Do not leak secrets or private logs.

Concrete Crewrift commands, Notsus function ownership, movement/voting/chat role details, and Crewrift replay debugging live only in this tier.

Coworld player optimization loop, artifact routing, local-vs-hosted evaluation, baseline/champion comparison, and optimizer workbench semantics live here.

Evidence-first debugging, smallest-owner changes, narrow regressions, before/after comparison, and provenance hygiene survive outside Coworlds.

Dropped: current extraction transcript, off-topic Gastown/Beads/Dolt/boot/disk sessions, raw source copies, and unsupported performance claims.

METADATA.md
haul/
universal/
crewrift-specific/
optimization-loop-specific/
generic/
null-tier.md
review.html