IN‑FLIGHT · OPO training & simulation Flight Deck : SENTINEL HOTWASH See it: THE CABIN

IN‑FLIGHT

A flight simulator for the hardest conversations in medicine — staffed by the first synthetic counterpart that is still there while you talk.

A resident-mind simulation program for organ procurement organizations. Every high-stakes conversation in the building — the authorization, the referral call, the offer, the huddle — rehearsable at 2 a.m., against a family that reacts while the words are still in the air. Built on the same open substrate as the rest of this site.

Proposal · pre-build Unreal Engine · MetaHuman Runs on a resident local mind Zero-egress mode Ships with its tape
Named twice

At the organ procurement organization where this program was born, the main conference room — the one where every hard decision in the building eventually lands — is called FLIGHT. This page is a tribute paid forward.

And in the craft itself: in-flight is where a resident mind lives. It reads the room while a sentence is still moving, holds silence as a decision, and can change course mid-utterance — un-saying a line it has already begun, because it knows exactly which words reached the air. The whole program is named for the moment a sentence turns midair.


§ 01The job

Aviation rehearses its rarest moments.
Procurement performs them live.

A century ago aviation reached a conclusion the rest of the world is still catching up to: the rare, catastrophic, seconds-wide moment is precisely the one you rehearse — at zero cost of failure, on demand, at night, as many times as it takes. The full-motion simulator is why an airline pilot's first engine-out is never their first engine-out.

Medicine adopted the simulator — for the body. Mannequins arrest, bleed, and code on schedule, and clinical skills are drilled on them every day. But the organ procurement organization's rarest and highest-leverage moments are not procedures. They are conversations — under time, under grief, under policy, where how it is said is the outcome. And the conversation never got a simulator, for one reason: a conversation needs a counterpart with a mind, and there was no mind to install. The only rehearsal partner money could buy was another human being — an actor, a colleague, a trainer reading from a script. Scheduled. Expensive. Daytime. And everyone in the room knows it's Carol from education.

So the rarest skill in the building is still trained the way aviation trained before simulators: live, on the real thing, with everything at stake.

Family authorization is the deepest room — and it is one room on a long corridor.
The deepest door

The Family Room

Authorization and first-person disclosure — one trained person, a family in the worst hours of their lives, and a conversation where every word, pause, and second of timing carries weight. The founding module, with its own full specification.

The clock starts here

The Referral Line

The call from the hospital unit. Triage under minutes: clinical intake, ventilator status, the questions that must be asked and the ones that must wait — on the phone, at speed, with a nurse who has nine other patients.

2 a.m. · pushback

The Offer Call

A surgeon on the line, a clock on the organ, and hard questions arriving faster than the chart can answer them. Holding your ground — accurately, calmly, at 2 a.m. — is a rehearsable skill that has never had a rehearsal room.

Before the approach

The Huddle

The bedside team, the chaplain, the timing conversation with hospital staff — the coordination that decides whether the family ever meets the right person at the right moment.

Onboarding

The First Solo

New-coordinator training: the first unaccompanied call, taken in a simulator instead of on a stranger's worst night. The growth tape shows the education department exactly what six weeks of practice changed.

Already on this site

The Board & the Handoff

The working floor's other seams already have residents: ROUNDS watches the board, VIGIL watches the sign-out. IN-FLIGHT is the same substrate, pointed at rehearsal instead of production.

Different rooms, different scrubs — one skill family: a high-stakes conversation under a clock, where the counterpart's face changes while you speak, where interruption is the terrain, and where the craft's hardest move is not filling a silence. Train the family, not just the flagship scene.


§ 02The gauntlet

Five generations of simulators.
One hole, faithfully preserved.

Field notes, not conjecture — every generation below was evaluated in the course of real work inside one OPO's education program. Each bought a new periphery. Each kept the same hole at the center.

GenerationWhat it isWhy it fails this roomMissing organ
Mannequin-class
SimMan-era
Instrumented physical manikin, control console, dedicated sim lab Six figures plus a full-time trainer puppeteering it; built for clinical skills, not conversation; and its datastreams still need a human to make sense of them — the instrument records, but nothing in it reads Everything above the airway
Packaged-VR-class
SimX-era
Multiplayer VR scenarios sold bundled with their own headsets No language model anywhere in the product — a staff member voices the patient. A puppet show with better goggles, at the staffing cost of the mannequin. And no OPO module exists: the vertical is too niche for a horizontal catalog A mind, period
In-engine NPC-class
Convai-era
Turn-based LLM characters inside Unreal Finally in the engine — and turn-based to a fault: push a key to talk. The character emotes only while speaking and is absent while the trainee talks. Non-technical staff spontaneously noticed the delay and called it unreal Presence
Video-avatar-class
Tavus-era
Conversational 2D video personas The fastest rented turn — and flat. A face on a screen is a kiosk: it cannot enter the engine, the captured room, or a headset. The scene, the props, the second family member — none of it exists Embodiment
Voice-model-class
CSM-era
Expressive conversational speech models The best mouth in the industry — no face, no body, no room. And beneath the prosody, the language backbone still takes turns: the lowest layer of the stack is request-scoped by its own documentation Everything but the mouth

Read the pattern down the last column. Each generation is a real advance in periphery — better body, better goggles, better engine, better face, better mouth — wrapped around the same vacancy. The hole was never intelligence; rented intelligence is superb and getting cheaper by the quarter. The hole is presence: a mind that exists while the trainee is talking — which is when a grieving family exists most.

The second lesson in row two — read it twice

The packaged-VR vendors will never build the OPO module. Fifty-five organizations is a rounding error in a horizontal catalog's market — and the entire world in yours. A vertical this deep gets served exactly one way: when the industry builds its own instrument.

That is not a complaint about vendors. It is the shape of the market, and it is the reason this page exists on this site — the one place where the policy corpus, the resident substrate, and the field knowledge already live under one roof.


§ 03The missing organ

Every generation made the turn faster.
None of them removed it.

Endpoint detectors got smarter. Inference reached thousands of tokens a second. None of it touches the defect, because speed only compresses the moment the machine exists — it shrinks the flash and enlarges the absence. The fix is not a faster turn. It is no turn: a resident mind, on local silicon, that ingests the scene as it streams and decides at every boundary whether this instant deserves a word, a glance, or nothing.

That primitive now exists, measured, on one desk in Dallas. It is called FUSOR — the resident-kernel substrate behind everything on this site — and in this room it buys exactly the five things no prior generation could:

The listening face

A mother's face changes at the wrong word as it is spoken — not 800 ms after the paragraph ends. The ingest side of the mind drives the face every frame the trainee talks: the brow off prediction error, the nod off boundary timing, the mouth off a valence probe. Reading the room in real time is the core skill being trained, and only a mind that is present in real time can teach it.

The third seat

Reality streams into the same mind: the monitor alarm, the phone, the chaplain at the door, the second unbuckled seatbelt of a scenario event. The world can take the floor mid-sentence — and distractor discipline, acknowledging without being derailed, becomes a trainable competency instead of an untestable virtue.

Silence, held

Families go quiet. The craft's hardest skill is not filling that silence. A turn-based system structurally cannot hold a meaningful ten-second silence — to it, silence is just waiting for input. To a resident, every silent boundary is a logged decision with a margin. Not speaking finally shows up on the record as the skill it is.

The turn midair

Commit-awareness: the mind knows exactly which words reached the air, so "I'm sorry — let me say that differently" is coherent, and the clumsy phrase from minute three can resurface at minute twenty as accrued distrust, not retrieved text. It can also un-say a line it has already begun. This is the mechanic the page is named for.

One trunk, many mouths

The spouse and the adult son who disagrees are branches of one mind — shared witness, divergent disposition. The son reacts to what was said to his mother because one mind heard it said. Turn-based multi-NPC scenes are strangers taking turns; a family is a single memory wearing several faces.

The substrate's plate Measured

These are the resident kernel's bench ratings — dated, receipted, from the reference build's public runs ledger. They are substrate numbers, not simulator numbers; the distinction is the contract in § 07.

QuantityRatingReceipt
Thought-boundary detection (judge exactly where a claim completes)0.97 vs 0.02ingest logits, free
Completed thought → judged~44 msone-token probe
Abort a forming utterance13 µsbranch kill
Fork a family member0 MiBbranch of one trunk
Three seats, one attention state1.208×worst-case co-decode
Honest recall across a working day≥ 98,304 tokmulti-hop spans
Tuned restraint at dial zero6.7% firereal-distribution soak
Live judgment on real streamsfirst hours bankedignition receipt, 2026-08-12

REFERENCE BENCH: ONE CONSUMER GPU · 9B-CLASS OPEN WEIGHTS · WINDOWS. SIMULATOR-SPECIFIC NUMBERS DO NOT EXIST YET AND THIS PAGE WILL NOT PRETEND THEY DO — TRANSFER OF THE TUNED DISPOSITION TO THIS SCENE IS A BET, AND § 07 NAMES ITS EXECUTIONERS.


§ 04Why now

Every organ arrived in the last eighteen months.
Separately. Shipped. Waiting for a center.

This program was not possible in 2024. It is barely, precisely possible now — and the assembly window opened one component at a time:

Sep 2025

The speaking face went open. NVIDIA open-sourced Audio2Face — photoreal lip-sync and facial performance from audio, running locally, no cloud seat. Shipped

2024 →

The body from sparse intent. MaskedMimic-class physics trackers synthesize full-body motion from sparse constraints — the mind emits meaning-rate intent tokens; the controller fills the hertz. The embrace and the walk to the door stop being animation problems. Shipped

2025–26

The room became a capture, not a purchase. Gaussian splatting went industry-wide — phone-grade pipelines on one end, and 360 cameras now shipping spatial capture to 3D scene as an in-app feature on the other. Splat your organization's own family consultation room, ICU corridor, quiet room — and the trainee rehearses where the real conversation happens. Reported · verify capture path at purchase

Jun 2026

The set builds itself. Unreal Engine 5.8 shipped an official MCP server — coding agents drive actors, materials, lighting, and tests from inside the editor. The scene is assembled by agents and reviewed by humans. Shipped · experimental

Aug 2026

The mind stepped a size class — this week. Qwen3.8-27B landed under Apache 2.0: a dense 27B with native vision, a 262,144-token context, and quantized builds already running in the local loaders. The reader the resident rides went from 9B to near-frontier on a single prosumer card — same loop, same laws, a far deeper mind, and the family scene is a vision problem as much as a language one. Shipped

Standing

The council is cheap. Wholesale-speed serving (Cerebras-class) drafts in-scene dialogue at thousands of tokens a second; frontier deep tiers (Fable / Opus-class) grade the tape and author the debrief between beats, never inside them. And in zero-egress mode every tier collapses onto the workstation: line quality degrades gracefully, and nothing — nothing — leaves the building. Shipped at each tier Collapse mode: budget

2026-08-12

The center organ has receipts. The resident kernel ran its first live hours on real streams — measured, ledgered, banked. The one organ every previous generation was missing is the one organ this stack now starts from. Measured

Five generations of simulators kept upgrading the periphery around the hole. In eighteen months, every periphery went commodity — and the hole grew a tenant.


§ 05The program

One substrate. Many rooms.
A curriculum that reads the tape.

IN-FLIGHT is not one scenario. It is a simulation program: one resident trunk, one council, one ledger — and modules that are nothing more than scenario families pointed at different doors on the corridor.

Founding module

The Family Room

Authorization & disclosure, one to three family members, faith and cultural casting, interpreter-mediated variants, the disagreeing son as a branch of the same trunk. Carries the full ethics plate below, and its own specification.

Module

The Referral Line

Voice-only — which makes it the cheapest module to build and the first to run (§ 09). Clinical triage intake against a charge nurse who is busy, distracted, and right.

Module

The Offer Call

Surgeon pushback under an organ clock. The scenario bank varies the case; the director varies the pressure; the tape shows exactly where composure held and where it slipped.

Module

The Huddle

Multi-party hospital-staff coordination — the approach-timing conversation where most authorization outcomes are quietly decided before any family is met.

Program spine

The Director

A hidden scenario engine allocates pressure — the chaplain arrives, the phone rings, the son pushes back now — under one cardinal rule: if the trainee would perceive the arrangement, it has failed. Emergence is legal for allocation, never adjudication: it schedules when you get tested; it never decides what the family feels.

Program spine

The Instrument

Every judgment, hold, and margin is ledgered. The debrief view lays trainee audio on one rail and the family's surprise-and-valence trace on another: the word where her trace spiked, the eight-second silence you held (margin shown), the near-rupture you never saw. No standardized-patient actor, at any price, hands the coach millisecond-resolved emotional telemetry. The assessment instrument is a side effect of the architecture.

Program spine

The Curriculum

Residual-directed: the director schedules the next scenario where the trainee's tape is weakest — struggles with silence get silence; struggles with the angry son get the son. Training by surprise: densify where the learner is thinnest, automatically, session over session, on a growth tape the education department can actually read.

Program spine

The Coach, promoted

No puppeteer, no scheduling an actor, no daytime-only. The trainer stops voicing the patient and becomes what they always should have been: coach of record, reviewing the instrument cut. And the room is open at 2 a.m. — because that is when the real calls come, and rehearsal should live where the job lives.

On headsets: the room is captured in 3D on day one, so VR is a viewport decision, not a rebuild — bring any headset, or none. No goggles are bundled with this program, on principle: the value is in the mind, not the periphery. Five generations of periphery-first products are how the industry got here.


§ 06The plate that cannot be unbolted

This system optimizes the conversation.
Never the outcome.

In this domain the ethics section is not an appendix; it is load-bearing structure, and it is enforced in the architecture rather than promised in a manual.

Non-negotiables · enforced by construction
  • The unbuildable failure mode. The simulated family's decision is not an available optimization target or score anywhere in the system — not in trainee metrics, not in the director's objective, not in the rubric. Fidelity to best practice only: compassion, accuracy, timing, informed and unpressured decision-making. A trainer that optimized for yes would be teaching coercion; this one cannot be configured to, because the target does not exist in the schema. The design goal is exactly one thing: no family ever says no because of a bad conversation. The decision remains the family's.
  • The trainee's tape is the trainee's. Sessions are emotionally exposed. Recordings belong to the person who made them; coach access is role-scoped; program metrics are de-identified aggregates. A training tool people fear becomes a tool people game.
  • Intensity is governed. A dial set at prebrief, a structured decompression after high-intensity scenes, and an always-available stop word that ends the scene without penalty or record beyond "session ended by trainee." This simulator will make people cry. That is by design — and it must be held by design.
  • Synthetic families only. Ever. The scenario bank is composite, built from published patterns and expert review — including cultural and faith review by people from the communities represented — and versioned, so a scenario that reads as caricature can be retired with a receipt. No real family's data, transcript, or likeness, under any circumstance.
  • Scope honesty. A training simulator — not a medical device, not a script generator for live conversations, and not a claim about authorization rates until the transfer study in § 07 runs for real. The system trains people. People talk to families.

§ 07The contract

Falsifiers, pre-registered.
The claims ship with their executioners.

Every vendor page in this industry is a promise. This one is a bet, and bets print their kill conditions. Each claim below releases with the build's tape and ledger — and each is structured to lose.

NameThe claimIt loses if
F-REALISMBlind expert requestors, shown sim sessions and human role-play recordings, cannot reliably rank the sim below the actors on believabilityThey can — the presence claim dies at the practitioner bar
F-LISTENING-FACEIn-speech facial response to semantic and prosodic events in < 400 ms; a turn-based comparison arm structurally scores zero while the trainee talksThe window is missed — the every-frame claim dies
F-DELAYThe delay that untrained staff spontaneously noticed in the turn-based generation is not reported in blinded sessions of this systemIt is — the core complaint survives the architecture change
F-SILENCEThe sim holds a directed ≥ 10-second meaningful silence, margins logged, without filling — and breaks it correctly when the moment turnsIt fills, or never breaks — the craft's hardest skill is not trainable here
F-MULTIPARTYThe second family member's reactions stay consistent with what was said to the first, across sessions, rated blindBranch incoherence — "one trunk, many mouths" is marketing
F-RUBRICDeep-tier rubric scores agree with expert human scorers at a pre-registered concordance barThey don't — grading stays fully human and the written debrief ships as notes, not scores
F-HOLDZero family interruptions on planted distractor events, all runsAny fire — the judgment is theater, and the honest component is a scene trigger
F-LOCALA full session runs on the reference workstation in zero-egress mode inside every latency budgetIt doesn't — the deployment story for regulated organizations dies until the hardware catches up
F-TRANSFERPre/post improvement on the organization's own validated communication rubric versus a shadowing-only control cohort, properly designedNo effect — the product is a vivid toy, and the next revision of this page says so
Not claimedMovement in real authorization rates is not asserted — not until F-TRANSFER exists as a real study, run with an education department as co-investigator. Anyone selling that number ahead of the study is selling.

§ 08See the mechanism

The demo is a cabin at altitude.
The destination is a family room.

Before any OPO installs anything, the mechanism itself will be public — as a five-minute scene anyone can watch: a flight attendant at 30,000 feet who reacts to a door event mid-word, runs both arms on the same tape, and ships with her falsifiers burned in.

And the mechanism is already running in this domain. Below is a recorded session from the referral-line module — not a mockup and not a storyboard: a real run of the resident kernel on a synthetic night-shift ICU, replayed from its own tape, with the identical script run a second time through a turn-based twin. Every number in it is read out of the tape at build time; both tapes are downloadable at the bottom of the block; and the limitations are printed inside it, including one pre-registered contrast that did not survive its control.

SYNTHETIC SCENARIO REAL RUN · every number below is read from the tape verifying…

02:14 — the referral line

A composite night-shift ICU. An OPO coordinator is on the speakerphone about bed 2; three staff share one attention state. The coordinator's twelve lines are hash-committed before the model loads, and the identical deck runs through both arms. No real patient, family, unit or organization is represented, and no PHI exists anywhere in this scenario or its tapes. The mind here plays the counterpart — it makes no clinical decision; it decides only whether this instant deserves a word.

29judgment occasions
8lines spoken
69holds, each on the tape
1sentence killed mid-word
0probe rounds, turn-based arm
0.0s hold emit abort
RESIDENT — three seats, one trunk, judging at every boundary
TURN-BASED TWIN — same deck, same seed, same weights; exists only at turns
What the tape says, computed. 29 judgment occasions; 8 lines spoken and 69 holds recorded — TECH never speaks at all. At the climax, BEDSIDE clears its dial mid-utterance at -0.24 and takes the floor; CHARGE's forming sentence is killed mid-word. The turn-based arm ran 0 probe rounds and produced 12 replies, one per coordinator line, re-reading 9,127 tokens of transcript to do it; 1 world event landed while it was mid-reply and is marked blind — a streaming reply can be stopped, but it cannot be informed.
Stated limitations, because they are on the tape too. The weights are a general 9B carrying a disposition tuned on desk streams and no training in this domain whatsoever. Two things follow, and this estate has already measured both. The register does not transfer: lines repeat (the near-duplicate damper caught 4 of them) and the payoff sentence is serviceable rather than good. And the disposition does not transfer either — an earlier bench in this same program (HOTWASH P0, 2026-08-16) measured this tune going silent on simulation-lab scenes, missing every operator-relevant moment, which is exactly why the seats here run on disclosed per-seat dials (CHARGE -1.30 · BEDSIDE -0.30 · TECH -0.20) rather than at zero. Those dials and the damper are interim scaffolding, not the tuned article; a domain-trained model is the fix, and it has not been trained.

And the one contrast this scene pre-registered did not survive its control. the same sentence renders at 104s and 249s (only the bed differs); in THIS talking run the margins move CHARGE +0.62 · BEDSIDE +0.79 · TECH +0.05. But the counterbalanced control says it is not the situation. Two cold runs of the same schedule — beds swapped, numbers swapped, dials set so nothing is ever spoken, the sampler consulted zero times in both — put the bed effect at +0.06 · +0.07 · -0.05 nats (CHARGE · BEDSIDE · TECH), against a scene-position component of +0.19 · +0.47 · -0.09; and the MAP 54-vs-78 effect at +0.15 · +0.26 · +0.09, against position +0.21 · +0.26 · +0.04. Read those against the ±0.25-nat band this bench treats as unresolvable. The separation visible in the talking run is the seats' own chatter raising the stakes as they speak — a feedback effect this estate has measured before — not the mind ranking one bed above another. So no ranking claim is made anywhere above, and the negative is printed here instead of omitted.
deck referral_line_v1 · deck sha256 19013f0707e34559eb1cbda7…
world sha256 edd218fd574ac6e67e87fb7a… · seed 17 · model Qwen3.5-9B-emit-v11-Q5_K_M.gguf · engine 99abe57-dirty
accelerated clock (17 ms of simulated time per generated token), so wall-latency claims are void here by our own rule; deck hashes are printed before the model loads (F-FRESH) and card text is tokenized only at fire time.
download the resident tape · download the twin tape — every number above recomputes from these files.
The public demo · same organ, one scene

THE CABIN — one scene, two seats, five minutes at altitude, and the first NPC who is still there while you talk. Same trunk, same laws, same tape discipline as everything on this page. Watch the sentence die midair; then imagine the room where that reflex matters most.

Board the Cabin →

And the simulator is not arriving alone. This site already carries the family it belongs to:

The knowledge layer

AORTA — the policy-fluent, confidence-calibrated mind: 21 chapters of OPTN policy, trained and citable. The framework that taught the model the job.

The presence layer

VIGIL and ROUNDS — residents watching the sign-out and the board, live, on the working floor. The same substrate, in production posture.

The rehearsal layer

IN-FLIGHT — the same substrate, pointed at training. The mind learned the policy; the floor got its watchmen; now the humans get a simulator.

The substrate

FUSOR-1 — the open resident kernel underneath all of it: the loop, the bus, and the laws. MIT. Measured. Receipted.


§ 09Status & build order

Proposal, pre-build —
and the first week is a kill test.

Every component named on this page exists, shipped, today. The stack as a whole is unproven — which is what the build order is for. It is sequenced so the transfer bet dies as early and as cheaply as it possibly can.

P0

The tabletop. Voice-only, no render: the resident harness re-aimed at one referral-line and one family scenario. Proves probe-on-delta, silence-holding, and interruption on this domain's distribution. One week class of work; kills or confirms the bet cheapest.

P1

One face. A single MetaHuman across the table, seated two-hander, listening lane live. F-LISTENING-FACE and F-DELAY run here.

P2

The family. Second branch — the disagreeing son — director nudges, the instrument cut, advisory rubric. F-MULTIPARTY, F-SILENCE, F-RUBRIC.

P3

The program. Scenario-bank governance, growth tapes, residual-directed curriculum — and the F-TRANSFER study designed with an education department as co-investigator.

P4

VR, if wanted. The room was captured in 3D on day one; the headset is a viewport decision, not a rebuild.

Bench: one 32 GB-class consumer-GPU workstation (mind + render, small scene), or the two-box topology — render box and mind box over LAN. Zero-egress is a first-class configuration, not a promise: grief conversations are sensitive by nature even when synthetic, and a CMS-regulated organization gets a switch.

STATUS: PROPOSAL · PRE-BUILD. SUBSTRATE FIGURES ARE MEASURED WITH DATED RECEIPTS; EVERYTHING ELSE ON THIS PAGE IS TAGGED SHIPPED, REPORTED, BUDGET, OR BET — AND THE BETS CARRY THEIR EXECUTIONERS IN § 07. THE BUILD RELEASES WITH ITS TAPE, ITS LEDGER, AND BOTH ARMS, OR IT DOES NOT RELEASE.

The rarest skill in the building, rehearsable at last — at 2 a.m., where the job lives. Aviation named the room fifty years ago: the simulator is where you meet the emergency before it meets you. Procurement finally gets one. It was always going to be called IN-FLIGHT.