Playtest ledger

Belief-changing runs get ink. One entry per real-model playtest that informed a design decision or a fun-test judgment; the book and learnings link here instead of restating numbers. Format: date · model · what was probed · what changed.

2026-07-08 · qwen2.5:14b (ollama) · first live fun-test probe — try-anything

Played over Tailscale on phone Safari against the real adjudicator. Fun-test under trial: "the world handed me a problem, I improvised from what's on my screen, and it judged me fairly — it said no when it should have."

What held (the seam works):

• "Smoke a fish" was refused because wood stock was actually 0 — a state-grounded no, exactly the earned refusal the prototype exists to prove.
• "Shout at the mountain" got a free, flavorful flat no — no AP charged, in-fiction narration.
• Illegal proposals bounced with reasons; nothing ever mutated state outside the delta vocabulary.

What broke (model-quality-bound, not architecture):

• Double-billing: spent wood and claimed the same wood as repair materials → wood burned, repair bounced. Produced the EXACTLY-ONCE prompt rule. See double-billing.
• Nonsense refusal: "melt snow for water" refused as "snow too clean to melt" — a wrong no, the mirror-image failure of yes-and.

Belief changed: the anti-yes-and architecture is doing its job; the residual failures track model quality. Recommendation recorded: run the verdict playtest on a Haiku-class model before judging the fun-test — requires ANTHROPIC_API_KEY on the desktop.

Verdict: not yet judged — deferred to a stronger adjudicator model.