菜单
Generate and test policy improvement hypotheses from eval evidence. Use when deciding what code or strategy change to make after analyzing replays and opponent behavior.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
| name | policy-hypothesis-loop |
| description | Generate and test policy improvement hypotheses from eval evidence. Use when deciding what code or strategy change to make after analyzing replays and opponent behavior. |
Make one clear change at a time unless running a deliberate candidate campaign.
Each hypothesis must include:
Example:
Observation: Crewmate mean is −8 driven by vote-timeout penalties in ~20% of meetings.
Causal guess: the cursor-confirm path misses the deadline when an LLM meeting call is slow.
Change: auto-submit the deterministic top-suspect vote at ≤72 ticks regardless of LLM state.
Expected movement: vote-timeout penalties → ~0; crewmate mean +8.
Risk: voting too early on weak evidence ejects crew.
Validation: 40 games vs broad field + 40 targeted; per-role + penalty decomposition.
Rollback: crewmate win rate or imposter win rate drops.
When testing several variants:
crewborg-optimization/crewborg-suspicion-tuning).map-navigation).Before changing policy, record the hypothesis. After eval, record:
Use to identify and download Coworld episode artifacts — replays, results, and per-agent logs — for completed episodes, or STREAMED live from a still-running experience request (--watch: each episode downloads as it finishes, no waiting for the batch). Triggers: 'download the replay/logs for episode X', 'pull the last N episodes <player> played', 'fetch artifacts for this experience request / pool / round', 'stream artifacts while the request runs', 'get the episode data for post-mortem'. Game-agnostic (Crewrift, amongthem, etc.).
Use to create and monitor Coworld experience requests — hosted batches of episodes you define (target, roster, roles, count) for evaluating a policy against a live field. Triggers: 'run crewborg vs the top opponents', 'make an experience request', 'request N hosted games', 'A/B a policy against the league', 'measure the imposter', 'set up an evaluation battery'. DEFAULT FLOW after create: launch the STREAMING pipeline in the background (step 4) — artifacts download and analysis prep overlap the still-running episodes; don't wait for the batch to drain.
Use to run your own built player policy in a LOCAL Coworld episode and watch it — the Gate-1 smoke test ('did my change take, does it connect → play → exit cleanly?'). Triggers: 'smoke test the player locally', 'run a local game', 'run my policy locally and watch it', 'does the build still work', 'open the replay locally'. Game-agnostic; this is correctness only, NOT a competitive matchup (you can't run other users' policies locally).
Use to submit an ALREADY-UPLOADED policy version to a Coworld league and monitor whether it QUALIFIES, then competes (and becomes champion). Triggers: 'submit crewborg to the league', 'did it qualify', 'is it disqualified / champion', 'watch the qualifier', 'monitor standings after submitting'. SUBMIT is the public, effectively-irreversible, champion-making action — explicit human go-ahead only. (Building + uploading a version is the separate `build & upload` skill.)
Use to decide whether a crewborg change ACTUALLY helped — A/B the candidate against the baseline head-to-head, fresh, right now, against the same field. Triggers: 'did my change help', 'compare v71 vs v70', 'A/B test crewborg', 'is the candidate better', 'did fixing the imposter regress the crew'. Runs two MATCHED fresh experience requests, diffs role-split metrics with significance (compare.py), renders a comparison report, and guides a qualitative side-by-side. It's also the 'designed run' that crewrift-experiment hands off to.
Use to build and query the Crewrift event warehouse — a policy-indexed DuckDB/Parquet dataset of per-tick gameplay events (kills, follows, proximity, votes, tasks, chat) over many episodes — for deep, mechanistic, cross-episode questions about a policy's behaviour. Triggers: 'how often is crewborg near a crewmate without killing', 'who trails crewborg', 'where does it abandon tasks', 'does it vote real imposters', 'build the warehouse for this round', 'query the events', 'stream an experience request into a warehouse'. For a FRESH experience request, stream_eval.py builds the warehouse WHILE the episodes run (the default; builds are incremental). This is the deep-dig mainstay after crewrift-survey flags something.