The owner's brief, 2026-07-31: 「think: what other levers we can pull to make the tool use more robust so that a new user gets a positive surprise 90% of the times. One possible lever: let the persona decide if he wants to act/speak…」 This page is the working plan and the scoreboard. The measure: the gauntlet's run-clean rate on the production shape — 26/32 ≈ 81% when the track opened.
| lever | what it does | measured | dial |
|---|---|---|---|
| The house's draw （server entropy, #11） | draw="words" — the house fills a deposit of your own; draw="pair" + a ?1/?1/?2 deck — the house draws the confusable pair at arm. The secret is chosen AFTER the speech published, by physics: nobody can leak what nobody chose, and no host converges on 斑马. The binding note and the dealer's map deliver the words; the history holds only the pool name. | hard leaks on the three leak scenarios 8 → 1 · the non-GM control（Newton）below-floor → 4/4 clean · the act call reached for it UNPROMPTED in self-play | lib/wordpools.py — always on |
| The second attempt | an act call whose EVERY form refused re-runs ONCE with the refusal reasons in front of it（「fix exactly what the refusal names — or kind=none」）; a missing on a deliberate turn retries as a decision. Never retried: a clean none, a partial arm, the gate. | converts the s43-r2 shape（a deckless deal → a whole round over a bare table）same-turn; both attempts billed; 3 smoke pins | always on |
| The silence door （the owner's lever, half 1） | on the floor producer's [MODE] hold the panel call is SKIPPED — structural, free silence; the humans' lines ride the next drain. Hard guards in code: ≥2 humans · never over a PROP verdict · never on a dispatch. Every hold books a coordinator event. | live probe: two humans bantering → hold, intent 「let the humans finish their thread」; the direct ask answered in character | silence_door · ON |
| The coda （the owner's lever, half 2） | the third fire of a GM's rhythm: explain → the cards land → ONE short line hands the first move（「Amy 先来」）. Only after a deliberate setup whose cards wait on players; skipped if anyone spoke meanwhile; its own act call pinned off. | live probe:「Cards are down… Amy, you go first.」— the full rhythm end to end; the exam's _turn now mirrors it | coda · ON |
The owner's brief: 「1) let the persona act/speak as he sees fit — the CC way … 2) study and reorg the FP/Prop/Act/Speech topology … 3) examples per tool … 4) play kits, same idea as skills.」 Built kits-first by the owner's priority call.
| move | what shipped | measured |
|---|---|---|
| 4 · Play kits （the kit shelf — its own page, with the five-beat flow diagram） | a game's researched RULEBOOK as a file（prototype/kits/, six kits: 卧底 · 大话骰 · 二十问 · 问答之夜 · 两真一假 · 真心话大冒险）: a HOST BRIEF the panel reads（appended once, cache-safe）+ a TABLE SPEC the act call reads while active. The shelf is a ninth tile on the tool sheet; a message naming a game gets a one-tap OFFER chip（server substring, no model call）. Variants are DIALS the host must declare — canonical rules researched, two profile errors corrected（1s wild by default in 大话骰 · spy-wins-at-three in 卧底）. | the control experiment: isaac-newton（zero GM training）hosted 卧底 BY THE BOOK with the kit — drawn pair, real ballot, whole pile turned at the end; the same persona without it pantomimed reveals, contradicted its own state and hand-wrote an unusable ballot |
| 1 · The loop （think→act→speak→verify） | the missing tail: deliberate_verify() — after phase 2's cards land, ONE look back（table vs speech, the landed handles in front of it）; expected verdict none, exceptional verdict ONE fix form through the same door, form-level capped, never re-arming a live card. The full turn is now prop（think）→ speech（explain）→ act（set）→ verify（check）→ coda（hand the move）— five beats, all behind the first bubble. | stubbed pins green（stash · none-verdict · consume-once · one-fix cap）; live run measured on the selfplay cells — the none verdict is the norm, the event ledger books every look |
| 3 · Examples | folded into the kits（per-game worked settings ARE the examples, loaded only when relevant）+ the act schema's existing draw/deck teachings — a standing examples block on every act call was rejected: the dilution law（s07: folding clauses into lists = 8/8→3/8）prices prompt weight paid on every turn, and a kit pays only when its game is on the table. | by design |
| 2 · Topology | studied, kept: prop（cheap think）→ FP ∥ act → speech → drain → phase 2 → verify → coda. The one merge considered — prop+FP as one call — was rejected: the prop's fail-open property is load-bearing（a dead judge must not silence staging）and the two run concurrently anyway, so the merge buys ~1s and spends a safety property. | decision recorded |
The question: can the persona OWN its game plan — write its own run sheet, check its own boxes — or does the world need a pointer? Replayed the three mary-test live failures on the SAME non-GM personas（Jensen's triple seal · Kim's frozen tally · Stan's 2T1F role flip）, three rounds:
| round | what changed | result |
|---|---|---|
| 1 — the sheet taught in prose alone | the kits' HOST BRIEF told the host to write a run sheet into its pad; the mirror surfaced settles + chip staleness | all three failures REPRODUCED, and run-sheet=no everywhere — the speech has no pad grammar under the split, and the hands（the act call）were never told. A teaching that reaches nobody who can write is not a teaching |
| 2 — identity keep-rules + the sheet reaches the hands | a re-declared chip LABEL keeps its address（a board rewrite can no longer reset a count）· one DRAWN envelope per author（a new draw replaces the old, close-with-release）· the kits' TABLE SPEC gains the pad step | nudge CLOSED（one envelope）· tally CLOSED（4 ticks, the sheet landed and updated）· flip still open — but the pad showed a CORRECT, self-updated plan: the failure had moved below the plan |
| 3 — one kit branch + one mirror fact | 2T1F's spec gains the host-as-storyteller branch（who="me"+my_answer — the hands had copied the literal who="uN"）· the mirror states every envelope's contents-state（rows in · answered by） | flip CLOSED: the host sealed his own lie at arm, revealed before the verdict, the sheet tracked the whole round. 3/3 |
| candidate | the idea | why it waits |
|---|---|---|
| verify-then-speak on ops | on OPERATE turns the speech runs parallel to the act and cannot know what landed — the C8 assembly fix narrows it, but a speech that READS the ledger before claiming would close claims_true structurally. | costs serializing the two calls（~3s latency）the parallel design bought on purpose; the claims seam is currently 4/4 green |
| the self-wake clock | a table armed and nobody moving for N seconds → the host nudges once（the clock instrument already exists as machinery）. | needs idle detection the worker doesn't have; the coda covers the commonest case（right after setup） |
| multi-fire beyond the coda | a persona loops speak/act freely inside one turn until done — the general form of the owner's lever. | unbounded cost and a new turn grammar; the deliberate turn + coda already give three fires on the beat that needs them |
| an EN word pool per room language | the pools carry zh + en; the room picks by its own language. A room in a third language falls back to zh. | third-language rooms are rare today; the fallback is safe |
The track's measure is the gauntlet, production shape（split · deliberate ON · N=4）. Opening state 2026-07-31 morning: 6/8 scenarios at floor, 26/32 runs clean, 9 hard leaks. After the day's levers: the three leak scenarios re-ran at 3/3 at floor, hard leaks 8 → 1, and the two mechanism defects the run surfaced（#26 the question into the void · #27 the half-move · #28 the graveyard channel）are closed with pins. The next full-battery run prices the 90% claim honestly — run it after the next change, not before.