| name | playtest |
| description | Play evolving-rpg in a repeatable, structured way and judge the results against the Covenant — use for any playtesting, balance check, or rule evaluation in this repo. |
Playtesting evolving-rpg
You are testing a game that evolves through ratified rules. Your job is to
produce evidence, judge it on two registers, and surface only direction and
polish questions to the human.
The procedure
- Baseline:
npm run play -- --policy all --seeds 12 --json. Record
escape/death rates, mean hp, dead-air per policy.
- Candidate: for any proposed rule,
npm run loop -- --seed 7 --rule r.json
(offline) or npm run loop -- --seed 7 (live model via dev server).
- Judge mechanically: the trial's verdict is demonstrated, not guessed —
but "sound" only means the exploiters could not break it. Look past it:
interactions with rules already in force, several modest rules compounding,
incentives that reshape play without breaking numbers.
- Judge thematically: the world is cold, quiet, attentive; second person;
concrete nouns; nothing shouts. A rule whose fiction contradicts its
mechanics fails even when every number passes.
- Deltas are the story: a sound rule moves at least one policy's
experience; a rule that moves nothing is noise; a rule that flips any
policy to unloseable or unwinnable is refused per Covenant M1.
Dispatch, don't do everything yourself
- Sweeps and reports →
playtester subagent (haiku).
- Rule verdicts that matter →
rules-warden subagent (sonnet).
- The session's strongest model reads their reports critically and owns the
call. A subagent's report is evidence, not a decision.
What reaches the human
Direction ("fighting still loses 90% — do we want combat viable?") and polish
("this speak-line breaks register"). Never mechanics you can verify yourself.
