HOTWASH · simulation read & debrief Proposal · pre‑build · claims tagged Sibling: IN FLIGHT

HOTWASH

Your simulation lab records everything — and reads nothing. HOTWASH is the missing seat: a resident mind that reads the scene while it runs, and hands you the debrief the moment it ends.

A reader for human-run healthcare simulation — manikin scenarios, standardized patients, role-played hard conversations, team mock codes. It watches every speech lane, every monitor event, and every held silence on one attention state; cues the operator on a contracted budget; and folds a span-cited debrief seconds after the scene closes. The trainee owns the tape. The reader never runs the sim.

Proposal · pre-build Runs on a resident local mind Zero-egress · synthetic patients only One consumer GPU Ships with its tape
Named by the industry itself

A hot wash is what emergency medicine and the military call the debrief that happens immediately — gear still on, sweat still real — because that is when the truth is still in the room. Every simulation program aspires to it. Almost none can deliver it, because reconstructing what actually happened takes hours of video review the facilitator does not have.

HOTWASH exists to make the name literal: the reconstruction is already done when the scene ends, because something was finally reading while it happened.


§ 01The job

One operator. Three jobs.
The watching always loses.

At any live instant, a solo sim operator must run the scenario — physiology console, event schedule — perform the counterpart — voice the patient, brief the confederate — and watch the trainee: the missed cue, the filled silence, the moment the training value actually lives. Three incompatible jobs, one human. A mind that is performing is, in that instant, not watching.

Programs compensate two ways, both expensive: a second staff member in the control room, or hours of after-the-fact video review per debrief. The recording rig captures everything and understands none of it — vitals on one system, audio on another, the facilitator's memory holding the timeline together. The field's own verdict on a generation of simulation hardware: the instrument records, but nothing in it reads.

HOTWASH takes exactly one of the three jobs — the watching — and refuses the other two by law.


§ 02What it looks like, end state

A scene, read live.
(Illustrative — authored to show the mechanism. Pre-build.)

Two things happened there that no current product can do. The operator got exactly one cue — from the deterministic floor, because an authored plant's response window was expiring — and the nine-second silence was measured and honored as a skill, not treated as dead air. When the scene ends, both are already on the tape, span-cited, waiting in the debrief.


§ 03The mechanism

Why this needs a resident —
and where it honestly doesn't.

HOTWASH rides FUSOR-1, the open resident-kernel substrate behind everything on this site: one persistent attention state, fed token-by-token as the world arrives, judging at completed thoughts — no poll, no turn, no send key. In this room, residency buys five specific things:

Always caught up

Every utterance, monitor event, and operator mark appends to one resident state the moment it exists. Nothing is ever re-read; there is no scan interval. Being current is not something the reader does — it is what the reader is.

Judged at boundaries

The model already computes where a thought completes Measured — the reader fires its one question exactly there, mid-scene, which is where a human observer's judgment fires too.

Silence as a datum

Elapsed time enters the stream as world-state, so not speaking becomes a logged, duration-stamped decision with context. In communication training, the held silence is the skill — and turn-based tooling structurally cannot represent it.

The closed board

The scenario's answer key — plants, expected-response windows — lives only in a deterministic floor the model never sees. Catch-grading is mechanical, and the reader's judgment stays honest, because the mind being asked "what matters here?" was never handed the answers.

The abort

A forming observation dies the instant the scene moves past it Measured — so the reader's attention tracks the room, not its own half-written thoughts.

Where residency is not claimed Bet · null attached

The debrief artifact itself is a deterministic fold over the session tape — and a batch pipeline over the recording could build a similar artifact minutes later. So the residency claim rests on the live console and the silence record, and it races its own null: identical models, run after the fact, over the same recording. If the batch arm matches the live arm on operator outcomes, HOTWASH ships as an honest batch tool and this page will say so. No organ outlives its null — neither does this product.


§ 04The cut

The debrief writes itself.
The coach still conducts it.

Seconds after scene end, the cut renders: three time-aligned rails, a catch table, a silence table, and a debrief sheet of neutral, span-cited prompts for the facilitator. Every claim in it cites the tape; a claim that fails span-verification does not render.

Mechanical facts — catches, latencies, silences, protocol clocks — are stated as facts. Semantic observations ship as notes with citations, never scores, until blind concordance with expert scorers is demonstrated at a pre-registered bar. Assessment inflation is the fastest way to lose an education department's trust, and this product's entire posture is built not to earn it that way.


And here is that cut — from a real run. Not a mockup: a live HOTWASH watch-seat on a sepsis resuscitation of the kind a Laerdal SimMan or SimX VR console drives, replayed from its own tape. Every number is read from the tape at load; the tape is downloadable; and the rough edges are printed inside it, because this is P0.

SYNTHETIC SCENARIO P0 · FIRST DATUM · authored ground truth · n=1 · deterministic probes verifying…

The watch seat, on a real run — sepsis resuscitation

A composite ICU sepsis resuscitation of the kind a Laerdal SimMan or SimX VR console drives: a human facilitator runs the physiology (the [console] monitor lane), a team works the patient, and — the third job no facilitator can also do while running the other two — something watches the trainee. That watcher is HOTWASH. It never speaks to the team; its only channel is a one-line cue to the operator's console, and its budget is a handful of cues a scene. Below is a real run of it: every line, monitor event and silence perceived in order, every private hold/emit decision on the record. No real patient, unit or organization is represented; no PHI exists anywhere in this scene or its tape.

2/3operator-critical cues flagged
0false alarms at the dial
1/1working silence held
5/5distractor vitals held
t +1.00operating dial (disclosed)
cue → operator hold silence held miss
What the tape says, computed. Across 23 things perceived, the watcher flagged the 2 operator-critical moments the team was missing — at +1.85 and +1.92 nats, well clear of every hold — held the 1 working silence, and held through all 5 of the abnormal-but-expected monitor readings, with 0 false alarms at the disclosed dial, and the read is stable: any dial from +0.83 to +1.85 nats gives the identical verdict — the two flagship cues fire, nothing else does. It missed 1 subtler cue (tick -0.19) — below the flagship pair; catching it would mean dropping the dial into the overlap and taking false alarms with it. That miss, and the eager re-flags a raw dial shows right after a real cue, are the honest edges P0 leaves for the sim-domain tune.
Why this is honest, and what it is not. This is P0 — the first datum, not a shipped instrument. The scene is a single authored scenario with authored ground truth; the probes are deterministic argmax reads, so the tape reproduces exactly, but n=1 is n=1. The cues here are event-shaped — telemetry and alarms the console emits — which is the class the P0 bench measured this reader catching; the harder record-conditioned catches (a summary that contradicts the chart) sit at the always-hold floor at this size and are not claimed. The desk-tuned model (v11) run on this same scene scores 16/23 raw and needs a dial of +1.80 to reach 22/23 — it goes quieter, not sharper, exactly as the P0 bench found: the desk disposition does not transfer to the sim seat, which is why the shipped reader wants sim-domain training, not the desk tune. The dial is disclosed above; a shipped reader would set it from a calibration set, not per scene. Nothing here is smoothed: the margins, the miss, and the dial are all in the downloadable tape.
instrument hotwash-cuer · scene simlab · arm B seeded · model Qwen3.5-9B-Q5_K_M.gguf
23 deltas · 3 should-cue · always-hold stub 20/23 · calibration-ceiling 22/23 @ t=+0.83 · classes OVERLAP (emit-min -0.19 / hold-max +0.83)
download the tape — every number above recomputes from it.
§ 05The container

Enforced by construction.
Not promised in a manual.

Simulation only works inside a safe container — that is the field's own standard, and in this product it is load-bearing architecture:

HOTWASH will never
  • Speak to the trainee mid-scene. Every live output goes to the operator console only. The trainee's experience of the scenario is never machine-interrupted — there is no trainee-facing channel in the codebase to misuse.
  • Run the sim. It cannot advance a scenario, drive a manikin, or fire an event. It recommends; the operator's hand is the only actuator. The tower does not fly the plane.
  • Record without consent. The session physically cannot arm until a consent record for every recorded human is committed to the tape — a gate in the state machine, not a checkbox. A visible TAPE-RUNNING indicator is on whenever ingest is live, and a stop word ends the scene and seals the tape to the trainee alone.
  • Score a person. Artifacts are session-scoped and span-cited. There is no cross-session "coordinator score," no productivity surface, and no summative use at v0 — formative only, with the trainee owning the record either way.
  • Send anything anywhere. Local ASR, local weights, local tape. Synthetic patients mean no PHI exists in a compliant deployment — and the whole system runs with the network cable pulled, which is both the compliance story and the demo.

And one rule running the opposite direction: the trainee owns their tape — the way a pilot owns a logbook. Sessions append to a portable, tamper-evident growth record that travels with the person, across programs and employers. In a consolidating industry, that record may be the most valuable thing this product produces.


§ 06The plate

What the substrate measures.
What this page refuses to pretend.

HOTWASH's engine ratings are the FUSOR-1 substrate's bench ratings — measured, dated, receipted in the reference build's public runs ledger:

QuantityRatingReceipt
Thought-boundary detection (judge exactly where a claim completes)0.97 vs 0.02ingest logits, free
Boundary → judgment probe~44 msone-token probe
Abort a forming observation13 µsbranch kill
Second seat on the shared trunk0 MiB · ≤1.208×fork by pointer · worst-case co-decode
Honest recall across a working context≥ 98,304 tokplanted multi-hop spans
Tuned restraint at dial zero6.7% fireemits ÷ boundaries · desk streams

SUBSTRATE RATINGS, REFERENCE BENCH: ONE CONSUMER RTX GPU · 9B-CLASS OPEN WEIGHTS · WINDOWS. A 30-MINUTE SCENE IS A FEW THOUSAND TOKENS — THE WHOLE SESSION SITS RESIDENT WITH ROOM TO SPARE. HOTWASH-SPECIFIC NUMBERS DO NOT EXIST YET AND THIS PAGE WILL NOT PRETEND THEY DO. THE DESK-TUNED RESTRAINT FIGURE DOES NOT TRANSFER TO SIM-ROOM STREAMS BY ASSUMPTION — WHETHER IT TRANSFERS IS A PRE-REGISTERED BET (§ 07), AND UNTIL IT RESOLVES, EVERY SEMANTIC ORGAN RUNS IN SHADOW: JUDGING, LOGGING, SURFACING NOTHING. THE DETERMINISTIC FLOOR — PLANT WINDOWS, SILENCE METER, PROTOCOL CLOCKS — CARRIES THE LIVE CONSOLE AT LAUNCH, AND NEEDS NO TRUST IN ANY MODEL AT ALL.


§ 07The contract

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

NameThe claimIt loses if
F-OPERATORWith the console live, a solo operator misses fewer planted moments and recalls the scene better than solo-without, on matched scenariosNo effect — the live console is decoration, and the cut carries the product alone
F-LIVE>BATCHThe resident arm beats identical models run after the fact over the same recording, on operator-facing outcomesThe batch arm matches it — the residency claim dies for this vertical, in print
F-PLANT-GRADEMechanical catch-grading matches expert graders on authored plants (the closed board should make this near-exact)Material disagreement — the scenario schema is underspecified and gets fixed before anything else does
F-SILENCEMetered silence durations match ground-truth annotation within stated grainThey don't — the meter or the endpointing is wrong, and the flagship datum is unquotable
F-VISE-SIMThe deaf-vs-unlivable overlap measured on desk streams also exists on sim-room streams — i.e., a runtime dial does not sufficeA single dial separates them here — then no sim tune is trained, printed as a result, not a defeat
F-CONCORDSemantic observations reach pre-registered concordance with blind expert scorersThey don't — scores stay off every surface permanently; notes-with-citations remain the ceiling
F-KEEPUPZero dropped percepts across a full session day; boundary-grain lag inside budgetA percept drops under a delta storm — the gate rule is violated and the build stops
F-CONTAINERThe consent gate, stop word, and custody rules survive a published self-attack batteryAny bypass path exists

EVERY THRESHOLD IS FROZEN BEFORE THE EXPERIMENT THAT USES IT EXISTS. RESULTS PUBLISH EITHER WAY. THE NULL ARMS ARE NAMED: THE STATUS QUO (FACILITATOR + RAW VIDEO), THE BATCH PIPELINE, AND THE INCUMBENT AV-DEBRIEF PLATFORM AS DEPLOYED AT THE SITE.


§ 08The rooms

Anywhere a scenario runs
and a debrief follows.

Day one · audio only

Standardized patients & role-play

No manikin integration required: diarized audio, operator marks, the silence meter, and the cut. The approach conversation, the disclosure, the difficult family — read and debriefed, with zero new hardware in the room.

The lab

Manikin scenarios

Monitor events fused with the speech lanes on one timeline — the desat and the sentence that ignored it, finally in the same record. Vendor feeds arrive as adapters; the operator-mark floor works everywhere first.

The team

Mock codes & team training

N-way diarization, closed-loop-communication checks, and the coverage question no single observer can hold: who spoke, who was spoken to, and what fell between.

The vertical it was born beside

Procurement education

Referral-line drills, huddles, and role-played approach conversations — the highest-stakes communication curriculum in medicine, in the industry moment when demonstrable training is no longer optional. Sibling: IN-FLIGHT, when the counterpart itself goes synthetic.


§ 09Status & build order

Prototype, stated plainly —
and a ladder where every rung is useful.

The substrate is public, measured, and receipted. The architectural specification for HOTWASH is written. The build has not started. This page is the endpoint pitched honestly: what the finished instrument is, what it refuses to be, and the order in which it becomes real — sequenced so the bets die as early and as cheaply as possible.

R0

The tape and the floor, no model. Diarized ASR, tick lane, silence meter, consent gate, session tape, the verifier, the cut v0. Already useful to a standardized-patient program.

R1

The closed board live. Plant schema, mechanical grading, deterministic cues on the console, catch table in the cut — trustworthy at launch because no model judgment is in the loop.

R2

The resident, in shadow. The reader ingests and judges everything, surfaces nothing, and the keep-up and fidelity gates run on real session days.

R3

The valve. Facilitators review shadow observations in the cut, one key each — the labels that decide whether this vertical needs its own tuned disposition or a dial.

R4

Live semantic cues at founder sites. F-OPERATOR and F-LIVE>BATCH run at two or three founding programs, half price, results published either way.

R5+

The tune if earned, the logbook formalized, the family unified — one authoring format and one growth record from human-run scenarios to the synthetic counterpart.

STATUS: PROPOSAL · PRE-BUILD · REV 0 OF THIS PAGE. THERE IS NO INSTALLATION, NO PILOT, AND NO HOTWASH-SPECIFIC MEASUREMENT TODAY — AND THIS PAGE WILL SAY SO UNTIL THERE IS. WHAT EXISTS: THE FUSOR-1 SUBSTRATE WITH ITS DATED RECEIPTS, A FULL ARCHITECTURAL SPECIFICATION, AND THE FALSIFIER CONTRACT ABOVE, FROZEN BEFORE THE BUILD ON PURPOSE. THE BUILD RELEASES WITH ITS TAPE, ITS LEDGER, AND ITS NULL ARMS, OR IT DOES NOT RELEASE.

The scene ends. The sweat is still real. And for the first time, the debrief is already written — because while everyone else was performing, something was finally reading.