The game device — the flow the server runs, the voice the persona keeps

D1 + D2 LIVE · 2026-08-02. The owner's frame, arrived at together after two sim cycles: the GM holds a「谁是卧底」app on a handheld device and hosts THROUGH it — like every real-world party app in the genre. The device deals, tracks the living, runs the ballot, buzzes the end; the human reads its screen aloud and keeps the table alive. Nobody calls that host incompetent for not counting votes in their head.

1 · The lesson this answers

Two sim cycles (20Q closed; 卧底 at 16 sweeps) taught one asymmetry, measured over and over:

Every prompt-side lever was tried in escalation — the check moved to the front of the settle block, then carried the dealer's map inline, then stated the string-equality itself — and each step lowered the flap rate without ever reaching zero. A game has 5–10 critical moments; even 95% per moment compounds into a broken game every few runs. A stochastic sampler cannot be prompted into an invariant — the prop-master study said it first;「the dependent variable is not an invariant」.

2 · The division of labor

3 · The kit becomes a cartridge

One file, three layers. The prose brief stays（the persona's understanding — tone, judgment rules, the spirit of the game）. A fenced flow block joins it（the device's program — validated against a closed schema whose only verbs are the eight instruments and four control forms）. The dials stay declarable.

== FLOW ==                                  # the device runs this; the persona never reads it
lobby:                                      # OWNER 2026-08-02: every cartridge gets this free
  join:   {ask: 这局谁玩？, min: 3, max: 9, close: 30s}
  #       roster ≤4 → OPT-OUT（all pre-joined, tap to sit out）· bigger rooms → opt-in taps
  #       under the floor → say so and fold back to chat; never deal a broken game
setup:
  deal:   {draw: pair, who: JOINERS, per: 1, visibility: own, dealer_sees: true, label: 谁是卧底}
  board:  {title: 谁是卧底, columns: 玩家|状态, counters: 轮次=1}
  moment: opening                           # persona: rules speech, in its own voice

round:
  await:  spoken_by_living                  # order-free; completeness is counted, not judged
  moment: recap                             # persona voices; the device supplies the verbatim list
  ballot: {who: living, options: living+弃权, close: 90s, blocking: host}
  on_tie: next_round                        # DIAL — declared, honored mechanically
  eliminate: top
  board:  update 状态                        # the row is the device's, not the persona's memory
  moment: result                            # facts handed over: the counts, the fallen
  end?:
    faces_identical → end(civilians)
    living <= 2     → end(spy)
    else            → round

end:
  reveal: all                               # the pile publishes — always, by the device
  moment: verdict                           # persona: who won and why, the goodbye, 再来一局?

abort:                                      # OWNER 2026-08-02: reachable from EVERY phase
  by:     player→table-vote(收/继续, 30s) | founder outright | idle 10m（the device folds itself）
  then:   reveal all → board wraps 中止 → back to chat
  #       the audit law holds even here: cards were dealt, so cards are shown; winner = none.
  #       the door is a UI TAP on the game's card — speech never steers, the persona RELAYS

dials:                                      # OWNER 2026-08-02: user flexibility, validated
  spies:            {by-players: 4-7→1 · 8-9→2}        # NOT dialable — rule integrity
  tie_rule:         {enum: no-elimination|revote, default: no-elimination}
  reveal_on_death:  {enum: never|role-only|full-card, default: never}   # the players' amendment
  #  set through TWO doors: ① chat —「这局死了要亮牌」— the PERSONA parses（a real use of
  #  the mind）and the setup moment returns dials as schema-forced fields; the device
  #  validates, out-of-range gets the default said aloud. ② UI — the lobby card shows
  #  declared dials as tappable settings. Undeclared = FIXED. Where nobody asks, the
  #  persona picks in character（the liveliness law）. A variant no cartridge declares →
  #  v1 says so honestly; D5 lets a user FORK the cartridge（edit → revalidate → their
  #  own shelf copy — the persona-clone pattern）. 20Q's sibling dial: questions 5..30.

The judgment moments carry the facts pre-chewed（the counts, the faces, the verbatim descriptions）— the persona reads its handheld's screen, exactly the settle-cue pattern that already works, minus every chance to act on the process.

The three moment kinds（owner, 2026-08-02 — improvised content is the persona's）

Trivia is the proof case: the device ships NO question bank — Hadid's quiz asks about cantilevers, MrBeast's about view counts, and the same cartridge runs both. Content is the persona's craft; commitments are the device's honesty. Where a kit's content is the HOUSE's instead（卧底's word pairs — entropy the persona must not know how to leak）, the flow says draw: pair and the device draws; the two sources are declared per slot, never mixed by accident.

# trivia（问答之夜）— the improvised-content shape
round:
  moment: compose {question → public, answer → seal}   # sealed first, posted second
  collect: {answers: all-players, close: 60s}           # the seal-collect the toolbox has
  moment: ruling per player {对/错/算对 + quip}          # judgment stays a mind's
  score:  board ++ from the ruling half
  end?:   rounds == N | score >= target → end

4 · The runner

• A small state machine per room — current phase, step cursor, the living set, the round chip — persisted in state.json beside the kit. Advanced by the events the world already emits（say/speak completeness, vote_result, deal settles, timers）: no new event types.
• It arms instruments directly. Process steps（the deal, the ballot over the living, the board row, the reveal）go through the same appliers the act call uses today — same events, same SSE, same UI. No act call, no prop verdict on process turns.
• Moments ride the wake door. A moment is a wake turn whose directive says what to voice and hands over the facts. The persona speaks; ops for color（appends, reacts）still pass; the flow does not wait on the words beyond its step timer.
• Humans keep every door. Players tap the same ballots and see the same cards; a human host could even play GM with the device exactly as the persona does.

5 · The liveliness law — MrBeast's game must not read like Hadid's

The owner's requirement, 2026-08-02: the flow is shared; the rhetoric must not be. The device's answer — it removes the parts where every host already sounded the same（the bookkeeping recitals, identical and error-prone）and returns the freed attention to the parts where they differ.

• Moment directives carry facts and obligations, never wording. 「The counts · the fallen · what must be known by the end of your speech」— style comes only from the profile. The FP firewall's law, applied to the device itself, with a hard lint: no example speech in any moment directive（one anchoring example converged a whole cast, measured in the G-audit era）.
• Everything between moments stays the persona's. Banter through the describes, teasing a vague answer, callbacks — free turns stay free; the device gates process, not speech.
• Personality picks the rules, safely. The dials become a structured SETUP moment: the device offers the enum（tie rule, turn order）, the persona chooses in character — Hadid's table plays strict canon, MrBeast's plays revote-drama — and the choice is mechanical forever after. Style decides the rules; the rules then never wobble.
• The freed context is the point. Six jobs in one forward pass become one: the whole profile + the room's history + the moment's facts serve VOICE alone. Liveliness should improve, not merely survive.
• Measured, never assumed — D2 acceptance gains a distinctness check: the same cartridge hosted by two personas, transcripts diffed by the judge. Two alike tables = a FAIL of this design, same severity as a broken ballot.

6 · What retires, what stays

The three tiers, top to bottom（owner-confirmed 2026-08-02）: ① a cartridge game — the device runs process, moments carry voice. ② a kit without a flow — today's classic path whole（prop → act ∥ speech, FP, wake, every cue and mask）. ③ no kit at all, or a game the shelf never heard of —「我们玩狼人杀」matches nothing, nothing auto-loads, and the persona hosts from its own memory, assembling raw instruments through the classic loop, every guardrail underneath. Improvisation is the fallback, never the ceiling: D5's compiler is the long-term door for unlisted games —「tell me your rules」→ a drafted cartridge → validated → played.

Cost: a settle moment today = prop + act + FP + cue machinery（~4 calls）; on the device it is one moment call — and most turns need none. The act-call cache lever precedent says measure it, not assume it.

7 · Authoring — where the stochastic step belongs

A user writes a rulebook in prose. A one-shot LLM call compiles it to the flow schema. The schema validator rejects anything outside the closed vocabulary; the sim harness dry-runs the cartridge; the auditor gates it. The sampling happens once, offline, where a human can review it — never on turn 47 of a live game. This is the user-generated-kits path the owner already named:「a kit is like pseudocode」— taken literally, compiled, not recited.

8 · Acceptance

• The auditor is the gate（exam/spy_audit.py, calibrated on the owner's own room judgments）: reveal-only-at-an-open-END · the spoken tally matches a real ballot · no mid-vote reveals · spontaneity · no card-less play. On the device, laws A/C/E hold by construction; the auditor proves it and watches B/D.
• The sims judge the voice — with process guaranteed, the troupe + judge measure what remains: does the table feel hosted?

9 · The build plan

10 · Ruled by the owner（2026-08-02）

• The lobby — every cartridge collects its players before dealing. Roster ≤4: opt-out（everyone pre-joined, tap to sit out）; bigger rooms: opt-in taps. Floor unmet → say so and fold back to chat. Everything downstream（the deal, the ballots' living, the completeness）keys on the JOINERS, never the room — spectators chat freely.
• The exit — an abort reachable from every phase, by UI tap（a player's tap → a quick 收/继续 table vote; the founder outright; a 10-minute idle fold）. An aborted game still publishes the pile — the audit law has no exceptions.
• The board is the device's screen. 状态, 轮次 and every mechanical row are the runner's writes; the GM reads the board aloud and never writes it — the entire row-lag saga was the GM failing as a scorekeeper, and two writers on one panel is the clobber bug by construction. Content-bearing rows（20Q's ruling column）come from structured moments: the ruling rides as a schema-forced choice（是/不是/说不清/不适用）beside the free-text quip in one call — the device writes the row from the structured half, the room hears the voice. A game wanting a scribble surface uses the pad or a second board, never the game board.

11 · Ruled by the owner, round two（2026-08-02）

• The visible hand: silent.「GM holds the game device — a card dealt from the device IS a card dealt from the GM.」The runner stamps the host's name on its arms; the persona's narration is the announcement; provenance lives in the ledger for debugging only.
• Pace: parallel first. Steps marked stage: after-speech（reveal-class）wait for the moment's own speak event — or its ~10s cap — then arm. Declared per step type in the schema; no runtime judgment.
• Moment model: current tier first, then A/B cheaper. v4-pro carries the voice at launch; the flash split is measured with the distinctness judge, never assumed（the flash-0731 lesson stands）.
• Soft rulings: the persona has the power — that's the fun. A too-specific description may be ruled dead: a RULING moment（forced 死/活 choice + the quip）whose consequence the device executes（eliminate → END check）. The host adjudicates, as real party hosts do; the table sees a judgment call, not a glitch.

12 · The call economy — who arms, and what it costs

The code arms everything in the flow. The runner drives the same appliers the act call uses today — same events, same cards, the GM's name on each — and the per-game LLM bill collapses to voice alone:

The pipeline, piece by piece: wake stays and is promoted — moments ARE wake turns with a directive and facts. prop effectively retires in device rooms. fp stays for staging（verdict-void in force; a slimming candidate, by measurement）. act is gone as a phase. speech becomes the moment call — the turn split collapses back to ONE call on device rooms: the split existed because one pass could not be a form-filler and a conversationalist at once, and with process in code the only form left is a four-token choice riding along. The classic path stays whole for kit-less rooms — dual-mode, unchanged.

13 · What the build taught（2026-08-02, D1 + D2）

The design survived contact. Everything §2–§6 claims — process in code, moments for voice, dual mode — went in as written, and the two defects that got through were both the same shape: a gate placed where the game runs instead of where the cartridge sits.

The screen is the piece the design did not name. §2 gave the persona facts at each moment and left the rest to voice — but a host also takes ordinary turns, and on those it knew nothing. A real host holding a device never has that problem: the screen is always there. One block, facts only, above everything: what round, who is in, who still owes a line, what the device is waiting for.

The instruments this phase added

• exam/device_dryrun.py — a whole game per scenario against the runner with no server, no API and no persona. 75 checks: both END doors, 3/4/5/8 seats, two spies, the tie, the revote dial, the dead man（including a stale-token no-op）, the two dead men racing, a dead player at both doors, a restart mid-game, the floor, the abort. It is where a process defect should die: here it costs a second and no tokens, while a live sim costs eight minutes and a bill. Gated in the smoketest.
• lib/gameflow.py -- lints the whole shelf: python lib/gameflow.py.
• live_sim.py spyre — the rematch: one table, two games, the second asked for in chat.
• live_sim.py spydiff — the distinctness check §5 demands: the same cartridge hosted by two personas, the transcripts diffed by a judge.
• live_sim.py spyadv — the table trying to break the device, six ways.
• live_sim.py spy8 · spysilent — eight seats（the two-spy band）and the four-minute dead man. The troupe grew from four seats to eight for the first.
• spy_audit.py learned what a device room is. Law D asked「did an urge open this ballot?」— on the device an urge CANNOT, so the question was noise. It now checks the device's OWN rule: every living player must have spoken since the last result. A stronger claim, not a waived one. Law A gained the abort as a lawful third END door. Both carry negative controls in the smoketest: a law that cannot fail is a law nobody is checking.

Acceptance, as measured

Open, and honest about it

• The lobby is opt-out only. §10's ruling has two halves; the built half seats every human at kit-load（correct for ≤4, which is every table tested）. The opt-in taps for bigger rooms need UI and are not built.
• The abort has no UI tap. The server door exists and every path publishes the pile — taking the cartridge off the table aborts, a runaway program folds itself, two silent rounds fold. The player-tap 收/继续 vote is not built.
• The ruling and compose moment kinds are schema-shaped but unexercised. 卧底 needs neither; D3's 20Q is what will prove them.
• The cost claim is unmeasured. §12 predicts the per-game bill collapses. The runs support it in shape（zero prop calls, zero act calls on device turns）but nobody has put a number on it yet.