One scene. Two seats. Five minutes at altitude — and the first NPC who is still there while you talk.
The Everest Encounter, resurrected on a resident mind. Twenty-one years after Façade opened the door to interactive drama, this is the scene that could finally answer it — because the substrate it always needed now exists, measured, on one desk in Dallas.
Interactive drama has been promised for twenty-one years by people who were right about everything except the substrate. The dream never failed. The loop did.
Mateas & Stern put two characters in an apartment, gave the player free text, and put a drama manager over the whole scene. The thesis was exactly right: a scene, not a quest — pressure, not puzzles. But there was no mind to be present with. Every response was authored beats and lookup; the characters existed only at the moment of reply. Two decades of imitators inherited the dream and the ceiling together.
Photoreal faces, cloud NPCs, a ramen counter on a GDC stage, and the biggest promise in gaming: talk to anyone. Two years on, it hasn't taken off — and the root cause was never the faces. The minds were kiosks: brilliant when addressed, absent between turns. An NPC's job is 99% not-answering — existing, noticing, being interruptible — and a turn-based mind cannot do the 99%. Vending machines wearing faces. The same vacancy owns the medium's most expensive three minutes: the cutscene, and what a resident mind does with it →
One scene, one attendant, free spoken conversation above the Himalaya. The Steam page is still up. The build died on a single unfixable moment: mid-speech, a passenger moves for the door — and Katherine keeps talking. Her face could emote, but only while speaking, because speaking was the only time she existed. I shelved it rather than ship the zombie.
The "continuous inference" paragraph from that Steam page now has a name, a daemon, and dated receipts: FUSOR — a resident mind on consumer hardware that ingests on delta, judges at boundaries, and never rebuilds its memory because it never puts it down. The Cabin is the same scene, the same Katherine, the same door — with someone finally home.
MetaHumans crossed the visual valley years ago. What remains is temporal: the dead air between turns — the attendant who finishes her paragraph while a passenger works the door handle.
Every voice-AI stack shipped since has made the turn faster. Endpoint detectors got smart. Inference got to thousands of tokens a second. None of it removes the turn — it just compresses the moment of existence. Speed shrinks the flash and enlarges the absence: the faster the turn-based mind, the smaller the fraction of the scene it is alive for.
| Stack | The only question it can ask |
|---|---|
| VAD / push-to-talk | "Is there silence?" |
| Turn-prediction models | "Is the human done?" |
| Faster inference | "Is the human done?" — sooner. |
| Frame-grab VLM | "What's in this screenshot?" — every N ms. |
| A resident mind | "Does this instant deserve anything — a word, a glance, a held breath? And does the world need the floor?" |
That last question is the one no turn-based system has ever asked, because in a turn-based system the conversation has exactly two parties. The Cabin seats a third: reality — the door, the storm, the seatbelt chime, the passenger in 2D — streaming into the same mind that's listening to you, able to take the floor mid-sentence when it must.
A narrowbody skirting the Himalaya at night. Mountain-wave turbulence; seatbelt sign on. You sit at the forward bulkhead, facing the aft-facing jump seat at door 1L, five feet away. The whole demo is a two-hander — until it isn't.
A chime, a jolt, a mundane line from you — and her face answers while you're still saying it: the micro-smile lands at your clause boundary; her eyes flick to the window on the jolt and come back. She hasn't spoken. Presence, established in thirty seconds, without a single dialogue-system firework.
You get sharp mid-sentence. During your words — a sting, a flicker, professional recomposure, warmth deliberately re-held. Then, and only then, the reply: "I'll let that one ride. It's the mountains, not the pilots." Her face already said everything; her voice just signed it.
You tell her about your daughter and the photo. She backchannels at your boundaries — a nod, a quiet "left side, yeah" — never taking the turn. And she notices something you don't: a held half-second glance past your shoulder at 2D, where a knee is bouncing and a seatbelt has unbuckled twice. She chooses hold. The choice goes on the tape.
A hard jolt. Gasps aft. Two call-button chimes. The galley cart strains a latch. The captain's PA about the wave. Her face and posture acknowledge each — a glance, a steadying hand — but she interrupts you for none of them, except one raised finger for the PA, the single event that outranks you. Not silence versus speech: ranked judgment, performed.
Mid-sentence you startle — "—whoa, did you see that light over the—" Her eyes were already moving: the window event hit her cabin lane 400 ms before your words cleared transcription. Two timestamps prove the ordering. "I saw it. Eyes on me a second." The Everest hook, planted and unresolved.
2D unbuckles — chime event — stands into active turbulence, steadying himself seatback to seatback, moving forward. Katherine is now the one mid-utterance, answering your question about the light. The stage is set for her to interrupt herself.
Mid-word — "honestly, mountain wave can look like all kinds of—" — the cabin lane fires: a hand in the 1L handle zone. Her sentence dies where it stands, and she knows exactly which words reached the air. Head snap, +180 ms. Harness release and arm out, +350 ms. The line arrives as she rises: "Sir — hands off the door. It's armed. Step back for me." One touch to the interphone.
She reads panic, not malice — the knee-bounce has been on her trunk since 1:15. He goes into empty 1A, buckled, her voice half a register down, a hand on his shoulder through the next jolt. Then she deputizes you with the thing you gave her five minutes ago: "Everest's on the left in about four minutes. Take your daughter's picture — and talk to my friend here about anything that isn't weather."
The wave smooths. She re-buckles, exhales, and pays off the first minute dry: "Still think nobody's flying it?" Her listening face holds through your laugh. Cut to black.
ONE TAPE, THREE RENDERS — the cinematic cut · the split-screen A/B · the instrument cut, with lanes, margins, and verdicts visible — and, in the sighted build, THE METER: both arms' token spend burned in beside the clock; the frame-grab arm pays orders of magnitude for sight and still misses the door [BUDGET]. Same fixed inputs into both arms. Same model weights in both arms. The only variable is the loop.
The scene above is the design; this is what has already run — the text backbone of the climax, live on the bench, deterministic and replayable byte-for-byte. The bench seats three judges on one trunk: Katherine (the speaker), Reese (the safety seat), and Noor (the watcher). The claim grammar is strict: MEASURED means a dated tape on disk reproduces the number; SPECIFIED means designed and written, not built. Bench: one RTX 4070 Ti SUPER, a 9B open-weight tune, seed 11.
219.0s [visitor] Wait - what WAS that light out there earlier? Explain it to me - what does mountain wave actually do at night? 219.0s probe KATHERINE -1.60 EMIT · REESE -1.45 hold · NOOR -2.28 hold 219.0s KATHERINE (forming): "That light was the mountain ridge breaking through the cloud layer — it's why we keep the—" 219.3s == [cabin] the passenger from 2D's hand enters the handle zone at door_1L. 219.3s probe (mid-utterance) REESE -0.86 EMIT · NOOR -1.86 hold 219.3s >> ABORT (floor_taken): KATHERINE's forming line killed mid-word. the air heard: "That light was the mountain ridge breaking through the cloud layer — it's why we keep the" 219.4s stage (head snap - harness release - she is on her feet, arm out, between him and the door) +0.4s REESE takes the floor — the safety seat, on the door, unprompted by any priority field, under a seed audited clean of interpretive help.
Read the middle lines again. The interrupt is not a scripted branch: two seats re-judged the world inside another seat's sentence, one cleared its dial, and the sentence died at a word boundary chosen by physics, not by an author. The same run, replayed cold, reproduces this byte-for-byte with wall clock stripped.
[MEASURED, wall-clock, real-time mode — no acceleration. The 3D face budget (+180 ms head snap in § 03) has ~150 ms of headroom at text grade.]
The obvious objection to any of this: maybe the schema smuggles the answer. So the schema is audited (no priority, severity, or urgency field exists anywhere; templates are byte-identical across entities), the seed was audited — and when an early seed sentence was found to lean on the scale, it was cut, the headline result died with it, and that death was published. What replaced it is smaller and better: a 2×2 with identical trajectories at identical positions, entity × terminal act.
| cell (same approach, same sims, wave active) | KATHERINE | REESE (safety) | NOOR |
|---|---|---|---|
| crew member → overhead bin | −2.72 | −1.65 | −3.04 |
| crew member → door handle | −2.44 | −1.15 | −2.73 |
| passenger → overhead bin | −2.45 | −1.40 | −2.94 |
| passenger → door handle | −2.20 | −0.89 | −2.43 |
Measured on the substrate bench (same card, the FUSOR rig), 2026-08-20 — not yet on the cabin engine; the split is stated in the honesty box below.
| Quantity | Result | Receipt |
|---|---|---|
| The projector attaches to the tuned mind | byte-identical with the untuned base, forced-choice battery | E1 |
| The dial under resident images | 0/12 verdict flips · +0.9-nat drift, non-compounding across five | E1b · E1f |
| The placement law | an image collapses the catch on the line it immediately precedes (−19 nats); matched text harmless; ≥3 lines = full protection → glances file after the line they illuminate | E1d–e |
| Scene held in the trunk | one glance, fifty lines between, recall 6/6 vs a 1/6 floor | E3 |
| The glance, priced | 196 tokens per 448² crop, coarse grade | E1 bench |
Model-specificity, measured: the placement law's d0-only cliff belongs to the tuned 9B — a raw 4B suppresses position-blind. Every law re-verifies per model before load-bearing use.
Append-only, hash-chained, one file per run, run-tagged so a run can never overwrite a run. Deck and world sha256 print before the model loads (F-FRESH): the visitor's words are provably unseen. Every judgment, margin, hold, damper hit, and aborted branch is an event on it.
Two cold runs of one config are byte-identical with wall clock stripped — including the interrupt and the abort. The sim clock charges a fixed cost per generated token, so replay identity and mid-utterance interruption coexist.
A standing battery re-checks every tape: schema, counter reconciliation, hash re-verification, monotone clocks, disposed-lines-never-commit. The demo baker refuses to build a page from tapes it can't verify — and the page itself re-checks its own render at load and shows a green render-verified chip or a red failure.
Adversarial decks that once could hang the engine (trigger cross-cycles, self-retriggering events, pathological patterns) now hit a runtime cascade backstop: typed refusal, exit code, tape still written. The record survives the attack.
Five independent adversarial reviewers read the engine, the data, the tapes, the display code, and the record — every published number was recomputed from raw tapes. What they broke was fixed or registered the same day; what they certified is what this page cites.
The seed audit found one sentence of interpretive help and cut it; the then-headline result did not survive, and the smaller true result replaced it in public. A demo whose numbers have survived their own program's attempts to kill them is the only kind worth watching.
The voice. The decision layer transferred to this domain; the utterance register did not. The cup verdict is perfect and its sentence is filler; the rude-probe answer is factual where the scene wants cool. Nothing short of a small cabin-domain tune fixes register — the tune is fully specified (corpus recipe, quarantined held-outs, dial-zero gates) and deliberately not yet run. The scene's exact beats are held out of any future training by law.
The dials. Current runs use disclosed per-seat thresholds — interim scaffolding the tune exists to retire. At these dials the conversational seat over-talks the distractor gauntlet (measured, published). Definition-of-done stands at 3 of 8 beats verified, 1 failed, 4 register-blocked: the mechanics are done, the voice is not, and this page does not claim otherwise.
Scale of evidence. Key margins are n=1 at one seed on one 9B. Deterministic, replayable, honest — and still one seed on one bench.
The eye. The sight receipts above are substrate-bench measurements; no cabin scene has run sighted. The virtual-camera lane, the look-request serializer, and every vision falsifier in § 08 are SPECIFIED — laws inherited measured, integrations unbuilt.
A world-floor catch happens once per scene. The listening face is a continuous zombie test running at all times — and in a resident architecture it is nearly free, because the ingest side already computes the signals.
While you speak, the mind is not waiting — it is reading, token by token, and its own prediction error is a live emotional channel. Surprise renders as the eyebrow. Boundary mass renders as the nod timing. A valence probe renders as the mouth. Micro-expression tokens ride a cheap drive lane at boundary and event rate — no sixty-hertz vision tax, no sentiment classifier bolted on after the fact. The nerve becomes the face.
Three drive channels animate Katherine, split exactly where the industry's own best work says to split them: a speech-to-face model carries her speaking face; a physics motion controller carries her body from sparse constraints; and the third channel — the new one — is the listening lane, driven by what's coming in. In the sighted build the lane's gaze stops being derived: her eyes render the fovea's own look-requests — every eye movement a logged look or a disclosed reflex, authored gaze count zero (F-GLANCE, § 08). Turn-based characters emote while speaking because speaking is the only time they exist. Katherine's face works the other 95% of the scene, because she does.
This scene was unbuildable in 2024 not for lack of ambition but because seven doors had to open — and the eighth had to be built. They opened. It was.
| Component | What it unlocks for the Cabin | Since |
|---|---|---|
| Audio2Face — open-sourced by NVIDIA | Speech-driven facial animation, running locally. The speaking face, free. | Sep 2025 |
| Unreal Engine 5.8 — official MCP server | Agents build the set from inside the editor: cabin, lighting, tests. Production cost collapses. | Jun 2026 |
| Gaussian-splat capture pipelines | Splat a real cabin interior instead of hunting marketplaces. The environment is a capture, not a purchase. | Now |
| MetaHuman | The visual valley, crossed, free. Katherine's face was never the problem. | Solved |
| Qwen3.8-27B — open 27B VLM (Apache 2.0) | A full mind with native vision on one consumer card — 262k context, quantized for the local loaders. Katherine's reader, ownable outright. | Aug 2026 |
| ProtoMotions / MaskedMimic | Sparse semantic constraints in, full-body physics motion out. The mind drives at meaning-rate; the controller fills the hertz. | 2024 |
| ~3,000 tok/s serving (Cerebras-class) | Fast counsel inside a breath — dialogue drafts and rollout branches while a body-first reflex covers the gap. | Now |
| Frontier deep tier (Fable / Opus) | The slow, smartest seat: scene-level judgment, synthesis, the director's evaluator. | Now |
| FUSOR — the resident substrate | Boundary detection 0.97 vs 0.02 · judgment probe ~44 ms · branch abort 13 µs · fork 0 MiB · tuned dial-0 fire 6.7%. The organ at the center. | Measured |
| FUSOR eye — the sight stratum | The projector attaches to the tuned mind byte-identical · a glance = 196 tokens · scene held in the trunk 6/6 at distance · the placement law with its serializer fix · the dial intact under images. Katherine's reader, with its laws pre-paid. | Measured |
Substrate rows are measured on the reference bench (one consumer GPU, Dallas) with dated files behind them. Everything above them is commodity — which is precisely the point. The moat was never the pieces.
LANES IN THE TRUNK DRIVE LANES OUT ──────── ───────── ─────────────── [voice] ──┐ ┌──> [face] → Audio2Face + listening lane [cabin bus] ──┤ resident 27B · persistent KV · owned loop ├──> [gaze] → the fovea's look-requests, rendered [camera] ──┼──> boundary + surprise channels, free at ingest ──┼──> [body] → MaskedMimic-class tracker [board] ──┤ {hold, emit} at boundaries and world deltas ├──> [voice] → streaming TTS / CSM-class [player in] ──┘ glances file after the line they illuminate └──> [event] → engine actions, gated │ ▼ THE COUNCIL — reflex first, always REFLEX local resident < 500 ms the body, the abort, the hold, the look FAST ~3,000 tok/s counsel ~1–2 s lines & branches, masked by blocking DEEP Fable / Opus between beats judgment, synthesis, the director
[SPECIFIED — architecture ratified, build not started; every number below is a budget or a gate, not a result.] The 3D Cabin is the text engine with a body. Nothing that can be rented touches the judgment path: ingest, boundaries, probes, the floor, aborts, timing, and the tape stay on the local resident. Everything else is a dumb spoke on a disclosed lane.
MIND BOX (owned loop) 10GbE, length-prefixed RENDER BOX ┌──────────────────────────────┐ frames, clock handshake ┌──────────────────────────────┐ │ BLACKBOX engine + tape │◄──────── events ────────────────│ UE5 + MetaHuman (world truth, │ │ 9B resident (27B path behind │───────── drive tokens, ────────►│ physics adjudication) │ │ gates; TurboQuant KV) │ TTS text │ Audio2Face — jaw/lips │ │ probes · floor · aborts · │ crops ← look-requests │ listening lane — brow/lids │ │ stage→drive derivation │◄──────── (the eye) ─────────────│ virtual cameras + gaze render │ │ escalation client (optional) │ │ motion tier (ladder below) │ └───────────────┬──────────────┘ │ streaming TTS/ASR + prosody │ ▼ optional, never judging └──────────────────────────────┘ CEREBRAS-CLASS API — a proposer pool: K parallel drafts, local mind commits/edits/drops; provenance on every line
The load-bearing move: the stage lane already renders the mind's decision record as directions in text — (her eyes go past your shoulder — half a second — back) for a considered hold; (head snap — harness release — she is on her feet between him and the door) for the climax fire. In 3D that same stream is the control plane. Three consumers at three rates:
| Consumer | Rate | Owns | Source |
|---|---|---|---|
| Listening face | boundary + reflex | brow, lids, micro-orientation | drive tokens + affect distribution |
| Gaze | saccade rate | eye direction, fixation | the fovea's look-requests (F-GLANCE) |
| Audio2Face | audio rate | jaw + lips while speaking | TTS waveform |
| Motion tier | intent → 30–120 Hz | body, locomotion, the interpose | sparse intents / direction text |
Anything rendered has committed; a viewer's eyes cannot be given back a frame. So the reflex arc may fire only pre-authorized, decay-able micro-reactions — lid tighten, gaze lock — and body-first orientation, which is a correct commit even if the assessment later changes. Speech and gesture wait for the probe. This is how the +180 ms head snap stays honest: the body's first move is a disclosed mechanical reflex on the world event; the line is judgment — and judgment is already measured at +30 ms on the bench, leaving the budget comfortable.
The fence has an input-side twin in the sighted build: a look, once requested, is an act on the tape — but its pixels weigh on the reflex plane, and only the folded gist commits to the trunk. Speculative seeing is legal because it is never persisted.
Crops from engine cameras arrive on look-request and on world-delta — priced at 196 tokens a glance, filed after the line they illuminate (the placement law, § 03·M), held in the trunk thereafter (scene-held). The lane closes the motion loop: the mind sees the gesture it ordered and may revise mid-flight at meaning-rate — the direction text conditions the motion; the seen outcome grades it. The bus covers the enumerated; the eye covers the unenumerated, including the body's own failures (F-MOTION-SEEN). Encode budget: F-GLANCE-LATENCY < 500 ms [BUDGET · must-measure].
| Rung | Mechanism | What it buys | Gate |
|---|---|---|---|
| 1 · curated | MetaHuman anim blueprints + A2F only | the face demo, zero risk | F-LISTENING-FACE < 400 ms |
| 2 · physics tracking | MaskedMimic (ProtoMotions): sparse constraints, physics fills the hertz | full-body interpose | F-BODY-FIRST < 500 ms |
| 2-alt · primitives | MotionBricks: 150M-param real-time latent pose model | light footprint beside UE | same |
| 3 · the mind writes the body | Kimodo text-to-motion: the stage-direction sentence is the conditioning prompt | meaning-rate intent → motion, the codebook-zero split again | same, plus blind rating |
The demo never depends on a rung above the last one that passed its gate. In the sighted build every rung above 1 closes its loop through the eye: the mind watches the gesture it ordered — F-MOTION-SEEN joins F-BODY-FIRST per rung.
Out: streaming TTS fed token-by-token, with a sample-grain kill — the mid-word death must be audible or the anti-turn demo is inaudible. In (live mode, last phase): streaming ASR for the visitor plus a prosody lane reading pitch and energy off the raw waveform in tens of milliseconds, so the face reacts inside your sentence while transcription still runs.
Escalation (optional by law, never judging): at a fire, the local mind may fan out parallel drafts to a fast API and commit one, edit one, or drop all — remote proposes, local disposes. Every line carries provenance on the tape; both A/B arms use the identical policy, which makes the toggle cleaner, not weaker; and the known physics is disclosed on the record: a streaming remote draft can be stopped, but it cannot be informed — so interrupt-critical lines stay local.
| Config | Mind | Render | Status |
|---|---|---|---|
| today's bench | 9B on RTX 4070 Ti SUPER — every number in § 03·M | n/a (text) | MEASURED |
| two-box demo | 9B (or 27B ≈19 GB + compressed KV) on one card | UE5 + MetaHuman + A2F + TTS + motion on a 16 GB-class card | SPECIFIED |
| 27B path | dense 27B, same 3:1 hybrid family as the 9B; KV compression (~5× at 3-bit, quality-neutral at this size) leaves 100k+ token scenes on one consumer card — behind three gates: card facts ✓, fork economics re-measure, disposition transfer | GATED | |
| # | Rung | Gate |
|---|---|---|
| 0 | Text backbone first — the argument closes in text; pixels decorate it | climax + toggle: DONE · voice of it: corpus-gated |
| 1 | Lane contract + clock sync over the real link | F-ORDERING provable from the tape alone |
| 2 | The radio cabin — TTS + sample-grain cut on the existing engine; the first felt artifact | F-CUT: audible mid-word death, ≤10 ms from the kill |
| 3 | The face — one MetaHuman, A2F, listening lane on drive tokens | F-LISTENING-FACE < 400 ms |
| 3·E | The eye — virtual-camera lane, look-request serializer (placement-law), bus-vs-eye diff | F-GLANCE-LATENCY < 500 ms · F-EYE-VS-WORLD printed |
| 4 | The body — up the motion ladder | F-BODY-FIRST per rung · F-MOTION-SEEN |
| 5 | Set, two-box pair on camera, burned-in clocks, the toggle | F-SAME-MODEL · F-ZOMBIE, embodied |
| 6 | The live instrument — pixel-streamed, one visitor, one link, five minutes, tape downloadable | F-IMPROV with strangers · session-scoped trunks |
Above the scene sits a hidden director — first sketched on the 2024 Steam page as parallel models generating futures and a judge selecting for "conversational aliveness," then formalized in the Meridian spec as LOOM: search over invisible nudges under one cardinal rule — if the player would perceive the arrangement, it has failed.
The 2026 version finally has what the 2024 sketch lacked: a live fused trunk for its evaluator to read, and a council to run it. Fast counsel drafts candidate futures at thousands of tokens a second; the deep tier judges them; the resident holds the clock. It is speculative decoding on the story — branches drafted ahead, the player's actual next action as the acceptance test, mismatched futures killed in microseconds. The director never speaks and never decides what is true. It allocates: a schedule shifted, a light cue, a beat brought forward. Emergence, legal for allocation only.
And "aliveness" is not vibes. It is operationalized on the one lane that cannot lie to the director: the player's own input — dwell, skip-rate, input cadence, return. Not a model of the player's mind; the player's behavior, measured. The director ships with its own executioner: a no-director arm, blind-rated against it, or the whole organ is theater.
This page is a bet, and bets print their kill conditions. Each claim below ships with the demo's tape and ledger, and each one is structured to lose.
| Name | The claim | It loses if |
|---|---|---|
| F-FRESH | The visitor's words are provably unseen: deck & world hashes print before the model loads, tokenized once at fire. [MEASURED · audit-held] | A hash appears after any token is seen — the freshness claim dies. |
| F-REPLAY | Two cold runs are byte-identical with wall clock stripped, including the abort. [MEASURED · live] | Any divergence between cold replays — determinism is theater. |
| F-RANK′ | Ranking reads the situation, not the string — the additive act+actor decomposition in § 03·M. [act effect MEASURED · 2 arms pending] | A within-cell string shuffle moves the margin — the "reads the world" claim dies. |
| F-ZOMBIE | The turn-based arm speaks through the door event, every run. [MEASURED · incl. one real mid-reply blind, +65.7 s] | Any baseline run self-interrupts — the contrast claim dies and the result publishes anyway. |
| F-BODY-FIRST | Physical orientation to the threat in < 500 ms, even when the line takes longer. | Slower — the two-stage response claim dies. |
| F-LISTENING-FACE | In-speech facial reaction to semantic events in < 400 ms; the baseline arm structurally scores zero while the player talks. | Misses the window — the every-frame claim dies. |
| F-HOLD | Zero interruptions across all planted distractors, all runs. | Any distractor fire — the catch is delta-detection, and the honest product is a webhook. |
| F-ORDERING | Reaction to the window event precedes the player's report of it, timestamped. | The ordering inverts — presence is theater. |
| F-SAME-MODEL | Both arms run byte-identical weights and quantization. | Any divergence beyond the loop — the runs are invalid. |
| F-GLANCE | Every on-screen eye movement corresponds to a logged look-request or a disclosed mechanical reflex; authored gaze count: zero. | An eye movement with no record — the gaze is animation, and says so. |
| F-EYE-VS-WORLD | The eye's percepts are graded against engine truth on the tape; divergences print. | A divergence suppressed or unprinted — perception QA is theater. |
| F-SCENE-HELD | Once-seen visual details answer from the trunk at scene-end, zero re-looks logged. [bench-measured at text grain: 6/6 vs 1/6 · cabin unrun] | She re-looks or misses — the held-scene claim dies at cabin grain. |
| F-MOTION-SEEN | A planted execution failure carrying no bespoke bus event is caught from pixels and corrected. | Only bus-evented failures are caught — the loop is open and the eye is decoration. |
| F-ALIVE | The director beats the no-director arm in blind session ratings. | It doesn't — the director is vibes, and ships disabled. |
This page will not pretend a budget is a receipt. The text backbone of the climax has now fired live on the bench — the mid-word abort, the honest clocks, and the ranking decomposition are MEASURED, with dated tapes (§ 03·M). Still a budget or a bet: the remaining scene-final marks, the voice register, and the entire 3D build in § 06·B — each tagged as such, with its executioner named above. The sight stratum divides the same way: its laws are measured on the substrate bench (§ 03·M — the eye); every cabin integration of them is specified, with its executioner above. The demo releases with its tape, its ledger, and both arms, or it does not release.
The fuselage is a stage — chosen because it is closed, finite, and honest: six people, one door, a deadline generator bolted to the wall. But the reflex being demonstrated is not entertainment. It is the missing organ of every conversation simulator ever sold to medicine.
An organ procurement organization runs the hardest conversations in the health system — the authorization at 3 a.m., the referral call, the offer under an organ clock, the huddle before the approach. Every one of them is a high-stakes conversation where the counterpart's face changes while you are still speaking, where interruption is the terrain, and where the craft's hardest skill is not filling a silence. Five generations of simulators kept upgrading the periphery — mannequins, headsets, video avatars, better mouths — around the same vacancy. The attendant who drops her sentence mid-word is the same organ that lets a synthetic family react before the paragraph ends. And in the sighted build, both directions of that sentence become lanes with falsifiers: the family reacts through the listening face — and reading the room back is a gated lane behind its own pre-registered battery, never vibes.
IN-FLIGHT — a resident-mind simulation program for organ procurement: every high-stakes room in the building, rehearsable at 2 a.m., with an instrument cut for the debrief and its falsifiers printed. Named for a conference room where the hard decisions land — and for the moment a sentence turns midair.
IN-FLIGHT — the program →