What it is
Output policy checks for a role-play NPC.
What it does: Routes each player message with a Jev Choice, then checks each local-model line with four Jev Nouls before the player sees it, with one rewrite and a safe fallback.
Why it belongs in the showcase: A clear, reusable split: the local model writes, Jev judges, code decides, and a separate scorer measures the result.
Caveats: Teaching code; thresholds are starting points only.
Evidence
Public Apache-2.0 repo linked to the episode video. The game first stages facts so the local model only gets what the current stage unlocks. Jev then picks the player message type (quest, small talk, tangent, breaking the game) with a Choice, and four Nouls check each draft line; a failed line gets one rewrite with the reason, then a prewritten fallback. A 40-message playtest compares naive, staged and gated modes; the scorer never asks Jev, so Jev does not mark its own work.
Caveats
- The README states it is for education only, not hardened or tested with real players
- The README does not print the playtest totals; they are in the episode and in local runs
- Needs Ollama with a local model plus a TypeSafe API key
Unofficial case page. Stars / list position observed for heat; do not treat heat as a quality SLA. Review the upstream README before you depend on it.