Smart NOA web simulator

SIMULATION — fictional patient — educational use only. Not a medical device. Not medical advice.

An educational, label-calibrated anesthesia pharmacology simulator that runs entirely in the browser. You build a fictional adult patient, review a computed assessment, plan a general anaesthetic or MAC/procedural sedation with doses auto-filled from the drug library, run an illustrative monitor, and debrief the case. It is not a medical device, not clinical decision support, and is not connected to any monitor, pump or health record. Nothing entered leaves the page: there are no requests to other origins (the only same-origin read is SOURCES.md, for the model-assumptions text).

The simulator lives in the Smart-NOA-Controller repository and is hosted at /simulator/ on collinbgeorge.com.

Run it locally

ES modules do not load from file://, so serve the folder with any static server:

cd smart-noa-controller            # repository root
python3 -m http.server 8765
# open http://localhost:8765/simulator/

No build step, framework, CDN or external font is used. The page sets a Content-Security-Policy (script-src 'self', no inline scripts or handlers).

How it works

The Sources, Model card and About & limitations panels can be opened from every step.

Guided mode and the automatic demonstration

Both live on the Run screen ("Guided & automatic modes" panel), work for GA and MAC plans and all four example cases, and use the existing speed controls (1×/5×/20×/60×). They use the patient and plan already entered and start the case from 00:00. They do not change any dose, rate, source or model value: drugs are given only by your plan steps, at the offsets you set, through the engine's plan scheduler.

• Guided mode (phase-gated). Each phase runs automatically from your approved plan; you confirm every transition with a gate button, and no transition proceeds on time alone when a state condition exists. GA: Transfer to OR (pre-op holding: enabled 5 min after the premedication bolus has ended, with a countdown) → OR setup (2 min) and preoxygenation (FiO2 1.0) → Approve & start induction (enabled after 3 min of preoxygenation and ≥ 10 min after the premedication ended; bag-mask ventilation, then induction plan steps) → Intubate (secure airway), enabled only when intubating conditions are met (all induction steps given, processed-EEG depth index ≤ 60, TOF count ≤ 1) → Enter maintenance mode → Start surgery (incision): the pre-incision verification — airway secured and ventilated, processed-EEG index in the GA band 40–60 (below 40 only with your acknowledgement), analgesia given per plan, TOF count ≤ 2 when the plan has a neuromuscular blocker, antibiotic started within its window (required), and the WHO time out confirmed by you → End surgery → Enter emergence / extubation mode → Extubate, enabled only while sim.extubationReadiness() is met. MAC: Transfer to OR → Approve & start sedation (after OR setup; FiO2 0.4) → Start procedure (sedation target held, antibiotic if indicated, WHO time out) → End procedure → complete when the sedation level is minimal or awake.
• Run full case automatically. A demonstration that walks the same sequence; gates pass once their state conditions hold (e.g. 30 min of surgery, 20 min of procedure). It never confirms the WHO time out, never acknowledges an index below the GA band and never gives missing prophylaxis: it pauses and waits for you (buttons in the panel; they do not end the mode).
• Same actions as the manual buttons. Every step calls the shared functions in engine/case-actions.js that the manual phase, airway and extubation controls also call, so interlocks, alarms and the event log behave exactly as in a manual run.
• Extubation is never forced. Extubation happens only when sim.extubationReadiness() is met (TOFR ≥ 0.9, Thilen 2023, plus the illustrative criteria); the "Extubate anyway" override is never used. If readiness is not met within 30 min of emergence, the demonstration stops and lists the unmet criteria; guided mode lists them and keeps Extubate disabled (you can take over).
• Pauses. Both modes pause, stop the clock and say why (on screen and through the aria-live region) when an interlock hold latches or an unacknowledged HIGH-priority alarm is active or newly raised. They never release a hold or acknowledge an alarm: you decide (e.g. Acknowledge, Resume hold), then press Resume. For an interlock hold you may instead press Continue with hold latched (this hold only) — explicit and logged, never a default, never chosen by the runner; held infusions stay at zero, no alarm is acknowledged, and any new latch (another trigger or held drug, or the trigger returning after MAP ≥ 62 / HR ≥ 50) or new HIGH alarm pauses again; the choice ends when the hold is released (AUTORUN_CONFIG.CONTINUE_WITH_HOLD_OPTION, simulator convention, owner-approved). Medium and low alarms (MAP < 65, depth over the sedation target, dose-range notices) do not pause; they are shown and announced as usual.
• Manual control ends it. Pause, Resume and Take over manually are always available. Using any other control on the monitor (phase, airway, drug, fluid, complication, sensor fault, step or play button, restart) ends guided/automatic mode and is logged; the case then continues manually. Speed, alarm Acknowledge and Resume hold do not end it.
• Labelled and recorded. While a mode runs, a banner says it is a scripted educational sequence, not a controller or a clinical recommendation. Each step it takes is logged with the prefix AUTO: (log kind auto). The debrief and the downloaded session JSON (automaticMode, summary.automatic) record the mode, the steps (automatic or confirmed by you), pauses, and where and why it stopped.
• Limits. The conditions are simulator conventions applied to illustrative model outputs (processed-EEG depth index, sedation level, TOF); only TOFR 0.9 (Thilen 2023), the ASA continuum levels, the GA band (S52/S53), the WHO time out (S54), the antibiotic window (S46) and the config.yaml interlock thresholds are sourced. Every condition and timing is in one block, engine/autorun-config.js, with its unit, rationale and provenance; the Model card lists them. In the healthy-adult example (seed 1) the illustrative model's MAP falls below 60 during maintenance, so the demonstration pauses there until you respond. A sequence can only start from 00:00; after you take over it cannot be re-entered mid-case.

Why the timing

Every technique-template step carries a timing rationale (timingRationale: text + sources, or "simulator convention"), shown under the step in the plan editor ("Why this time"), in the phase-timeline tooltip and legend, in the parameter inventory (component template_timing) and in MODEL_CARD.md ("Timing rules and template timing"). All timing thresholds live in one block, engine/timing.js TIMING_CONFIG.

• Loading, then maintenance. A maintenance infusion starts when its loading dose ends: dexmedetomidine loading 1 mcg/kg over 10 min, then maintenance at +10:00 (label "followed by a maintenance infusion", S9); propofol MAC initiation 0.5 mg/kg over 5 min, then maintenance (S11); IV lidocaine loading over ~10 min, then the infusion (S1). buildPlan() and the plan editor schedule it; the editor warns if you make them overlap.
• Induction order. Fentanyl ~3 min before propofol so it approaches peak effect before laryngoscopy — a simulator convention anchored to the fentanyl label ("maximal analgesic and respiratory depressant effect may not be noted for several minutes", S13). Lidocaine pretreatment immediately before propofol (S11). Rocuronium after loss of consciousness, beyond the propofol label onset of 40 s (S11). Dexamethasone at induction and ondansetron at the end of surgery (PONV consensus, S2/S3). Cefazolin within 60 min before incision (Bratzler, S46). Sugammadex after the agents are off, dose compared with the TOF depth of block (S6).
• Pre-op holding. Midazolam premedication is given in pre-op holding about 5 min before transfer to the OR (owner-stated practice; label onset 3–5 min, S17). Induction is allowed no earlier than transfer interval + OR setup + preoxygenation (10 min) after the premedication ends; plan drugs never start while a premedication bolus is running.
• End of surgery. In the opioid-free template, dexmedetomidine stops 15 min before the planned end of surgery (simulator convention: no source gives the interval). The automatic demonstration knows its planned end (30 min of surgery); in guided and manual mode the stop runs when you press End surgery (or at emergence at the latest).
• Pharmacology waits for pharmacology. A rocuronium top-up waits for TOF count ≥ 3 (label: at 25 % recovery of control T1 = 3 twitches, S12); a repeat midazolam bolus waits 4 min (label: over ≥ 2 min, then ≥ 2 min to evaluate, S17). Planned steps are held with the reason shown; a manual dose is warned and logged.
• MAC: second sedative after midazolam. In the MAC propofol template and the MAC example plan, propofol starts 4 min after the midazolam bolus starts (midazolam over ≥ 2 min, then ≥ 2 min to evaluate its effect — label, S17; ASA 2018 moderate-sedation guideline: allow time between doses to assess each, S32). Changed from +2:00 in the November 2026 timing cross-check.
• Timing evidence. Every timed element was cross-checked against FDA labels and guidelines (DAS 2015, ASA 2022 airway, ASA 2018 sedation, Thilen 2023, PONV consensus, Bratzler 2013, Foo 2021, SCCM 2023 RSI and others) in TIMING_EVIDENCE.md: what each source says, agreement, classification, the action taken, and a clinician-review checklist (no clinician review recorded yet).
• Timing check. The plan screen, the guided/automatic panel and the debrief show a Timing check (each rule with pass/warning and its source or convention).
• Processed-EEG depth index. The depth number is a processed-EEG depth index (BIS-style, simulated) — an illustrative model output, not a BIS value (BIS is a trademark of its owner). The GA band 40–60 is the protocol target of the B-Unaware (S52) and BAG-RECALL (S53) trials, neither of which showed the BIS protocol superior for preventing awareness. In this illustrative model 1 MAC of a volatile agent often reads just below 40, so the pre-incision gate asks you to acknowledge an index below the band rather than blocking. Above the band there is an engine alarm (all run modes): see Final-check fixes below.

November 2026 additions

• Technique templates (data/templates.js, engine/templates.js). Templates name drugs, regimens and phase offsets only; every dose and rate is filled at runtime by recommend() (no hard-coded doses), and locked drugs are skipped or replaced by the template fallback (e.g. sevoflurane → propofol infusion for MH susceptibility). GA — opioid-based: midazolam (optional), fentanyl, lidocaine pretreatment (optional), propofol, rocuronium, sevoflurane, dexamethasone, ondansetron, sugammadex, lactated Ringer's maintenance, guideline prophylaxis. GA — opioid-free: dexmedetomidine, ketamine and IV lidocaine (Foo 2021) instead of an opioid; lidocaine pretreatment is omitted because an IV lidocaine infusion is planned; evidence note from the POFA trial (more hypoxaemia and severe bradycardia, less PONV). MAC — dexmedetomidine: label procedural-sedation dosing (1 mcg/kg over 10 min, 0.6 mcg/kg/h titrated 0.2–1; 0.5 mcg/kg loading > 65 y), minimal/moderate/deep target, optional midazolam/propofol supplementation; sedation beyond deep is general anaesthesia (use a GA template); with interlocks on the Smart NOA ceiling (0.5 mcg/kg/h) is pre-filled. MAC — propofol: the existing technique.
• Lidocaine pretreatment before propofol: 10 mg (propofol label: 1 mL of 1%), counted in the cumulative lidocaine total; a LOW alarm and plan warning when it overlaps (within 4 h) with IV lidocaine analgesic dosing (Foo 2021). Injection pain is a teaching note only: propofol emulsion pH is 7–8.5 (label), so the pain is not from an acidic solution.
• IV fluids: lactated Ringer's and 0.9% sodium chloride (label composition and cautions). The labels give no fixed dose, so the 100 mL/h default, the 1000 mL/h input limit and the 250/500 mL bolus sizes are simulator conventions. Maintenance continues across phases. Both fluids use the same illustrative volume model; electrolytes and acid-base are not modelled.
• Antibiotic prophylaxis (timing only): Bratzler 2013 — cefazolin 2 g (3 g ≥ 120 kg) started ≤ 60 min before incision, redose every 4 h; with a severe β-lactam allergy, clindamycin 900 mg (≤ 60 min, redose 6 h) or vancomycin 15 mg/kg (≤ 120 min). MEDIUM alarms: "Surgery started without antibiotic prophylaxis in the recommended window" and "Prophylaxis redose due". Manual mode: a checklist warning before Start surgery can be acknowledged (the alarm still sounds at incision). Guided/automatic modes (timing review): a dose started within the window is a required pre-incision item — the clinician gives it from the panel; neither mode gives or acknowledges it on its own. "Prophylaxis indicated" is a plan checkbox (default on for GA, off for MAC).

Advisory haemodynamic rescue (owner-approved, November 2026)

Advisory only — the simulator never gives a vasopressor by itself. When MAP falls below 65 mmHg (the existing MAP-low alarm threshold, reused), an "Advisory: haemodynamic rescue" card on the Run screen (and in the guided/automatic panel) suggests one label-sourced rescue bolus: phenylephrine 40 mcg (label rescue range 40–100 mcg, every 1–2 min, ≤ 200 mcg in all [S18]) when HR ≥ 60 /min, ephedrine 5 mg (5–10 mg, ≤ 50 mg [S19]) when HR < 60 /min. The agent-by-HR rule and its 60 /min threshold are a simulator convention (no verified label or guideline statement). Doses come from the drug library (recommend()), never from the advisory code. The suggestion is withheld while the repeat interval since the last vasopressor bolus runs ("Reassess — next dose allowed in mm:ss"; phenylephrine 60 s from the label, ephedrine 120 s convention) and once the cumulative label maximum is reached ("label maximum reached — consider other causes / escalate; not simulated"); also during a sensor fault and the anaphylaxis scenario (epinephrine stays a manual choice). A non-drug checklist (depth/volatile, fluids, surgical/positioning causes, sensor artefact) is always shown. MAP > 120 gives text only (review depth/analgesia, check the sensor) — no drug or dose. "Give suggested bolus" is the clinician's one-click approval (logged ADVISORY accepted); "Dismiss" is logged. Nothing is accepted automatically in manual, guided or automatic mode; the advisory never acknowledges alarms or releases interlock holds. The debrief lists every advisory shown/accepted/dismissed with times. Code: engine/rescue-config.js, engine/rescue.js, ui/advisory.js; tests tests/rescue-2026-11.test.mjs and the advisory block of tests/ui.smoke.mjs.

Guided/automatic fixes (November 2026). (1) When the mode pauses itself (HIGH alarm, interlock hold, clinician item) the panel re-renders at once: a paused banner gives the reason and the exact action ("Paused — HIGH alarm: … Review the patient, acknowledge the alarm (Alarms panel), then press Resume."), Resume is enabled and focused, and the pause is announced. (2) MAC sedation below the plan target (e.g. the obese-OSA example at the label lower-bound propofol rate reaches only minimal sedation): guided mode offers the clinician Titrate (one step within the label range of the running agent — propofol 25 → 50 mcg/kg/min [S11]; midazolam at its regimen dose after the label 4-min interval [S17]; logged as clinician-approved) or Acknowledge and start procedure at current level (logged). The full demonstration pauses there and never chooses. Example doses are unchanged. (3) Example-plan steps now show the same "Why this time" rationale as the matching template step, or "Example plan timing (simulator convention)".

Final-check fixes (owner-approved, November 2026)

• No maintenance hypnotic after a lockout. When a lockout removes the only GA maintenance hypnotic (e.g. sevoflurane, desflurane or isoflurane for an MH-susceptible patient; propofol for a propofol allergy — a generic check over plan steps × the engine's lockouts, engine/plan-safety.js), the plan editor shows a blocking warning "No maintenance hypnotic after lockout: is contraindicated (). Add an alternative (e.g. propofol infusion — TIVA) before approving.", Continue to run stays disabled, guided Approve & start induction stays disabled with the same text, and the automatic demonstration refuses to start. The plan is never changed automatically: the optional Add propofol TIVA maintenance button adds the propofol maintenance regimen from recommend() at maintenance +0:00 and its stop at emergence, each with a "Why this time" rationale and checked by the Timing check; the case log records it as the clinician's addition.
• Depth-high (awareness-risk) alarm. An engine alarm in every run mode (manual, guided, automatic): HIGH, "Processed-EEG index > 60 for ≥ 60 s during general anaesthesia — risk of awareness; check hypnotic delivery". Armed only in GA with the airway secured after loss of consciousness, in induction/maintenance (surgery and closing included), and not after every maintenance hypnotic has been stopped once surgery ended; never in MAC/sedation or emergence. It reuses the GA band upper bound (TIMING_CONFIG.GA_DEPTH_BAND, S52/S53); the 60-s sustain (DEPTH_HIGH_ALARM_SUSTAIN_SEC), the clearing value 58 (DEPTH_HIGH_ALARM_CLEAR) and the HIGH priority are simulator conventions (parameter inventory, MODEL_CARD, TIMING_EVIDENCE row 64). Guided/automatic modes pause for it and never acknowledge it. The simulator has no depth-sensor fault. None of the example plans raises it (4 plans × guided/full × seeds 1–20).
• Unusable readings. Log and alarm text print an unusable value as "unusable" (e.g. Invalid sensor reading (HR=unusable, MAP=118)), never "NaN" (display only; Python parity unchanged). While the HR, BP/MAP or SpO2 sensor is faulted its tile shows "—" with the accessible label "no valid reading (sensor fault)", trend charts show a gap, and the debrief's HR/MAP ranges and minimum SpO2 skip the fault gaps.

Depth-low advisory (text only; owner-approved, November 2026)

The illustrative depth model reads about 34–39 at label-default maintenance doses (the healthy sevoflurane GA example spends about two thirds of maintenance below 40). Instead of recalibrating the model, the owner chose a text-only "lighten anaesthetic" advisory. When the simulated processed-EEG index stays below the GA band lower bound (40; TIMING_CONFIG.GA_DEPTH_BAND, the B-Unaware/BAG-RECALL protocol target [S52][S53], reused) for ≥ 300 s (DEPTH_LOW_ADVISORY_SUSTAIN_SEC; 120 s in 05k, lengthened in 05l to limit alert fatigue) during GA maintenance after incision, with the airway secured and after loss of consciousness, an "Advisory: depth index below the GA band" card appears on the Run screen and in the guided/automatic panel: "Depth index below the GA target band (40–60) for ≥ 120 s — the anaesthetic may be deeper than needed." It gives a checklist (review the volatile concentration — end-tidal / MAC — or the propofol infusion rate against the plan; consider haemodynamics — if hypotensive, depth may be contributing; confirm the planned depth target and surgical stimulus; check for sensor artefact) and the footnote "Advisory only — the simulator never changes a dose by itself. Simulated index (illustrative); the 40–60 band is the protocol target used in B-Unaware/BAG-RECALL [S52][S53]." It suggests no drug, dose, rate or percentage and makes no claim about outcomes. It ends when the index is ≥ 42 (DEPTH_LOW_ADVISORY_CLEAR) or at emergence / when every maintenance hypnotic is stopped after surgery; never in MAC/sedation or before loss of consciousness. Before incision the existing pre-incision item ("acknowledge index below the GA band") stays the gate. It is not an alarm: nothing is added to the alarm list, guided/automatic modes never pause for it, it never acknowledges anything. The only control is Dismiss (logged; hidden until the next episode). If the MAP < 65 advisory is shown at the same time, both cards stay visible and each names the other. The debrief lists every depth advisory shown/dismissed with times and the total time below the band. Sustain time and hysteresis are simulator conventions (parameter inventory, MODEL_CARD, TIMING_EVIDENCE row 65). On the example plans (seeds 1–20) it appears in every full automatic demonstration of the two sevoflurane GA examples and in 14/20 of the propofol-TIVA example, never in MAC. Code: engine/depth-advisory.js, ui/advisory.js (depthAdvisoryCard); tests tests/depth-advisory-2026-11.test.mjs and two blocks of tests/ui.smoke.mjs.

Palliative pain management and end-of-life care modules

Palliative / end-of-life simulation — fictional patient — educational only. Not medical advice. Palliative sedation is not euthanasia or medical aid in dying; laws differ by jurisdiction.

Choose Palliative pain management or End-of-life care with the module buttons under the header (the Perioperative simulator is unchanged). Each module has four steps — Case → Assess & plan → Run → Debrief — and an Ethics, law & communication panel that opens from every screen (definitions of palliative care, hospice, proportionate and continuous deep palliative sedation, euthanasia and MAID; intent, proportionality and the double-effect principle; capacity, surrogates, advance directives, interdisciplinary review, documentation, hydration, a jurisdiction note and paraphrased goals-of-care topics from the Serious Illness Conversation Guide, CC BY-NC-SA 4.0). The modules teach goals-of-care conversations, proportionality and double effect and make no outcome claims (no survival, life-span or efficacy claims). Excluded by design: any content, doses or guidance for medical aid in dying or euthanasia, or for hastening death.

• Cases (fictional): cancer bone pain at home hospice (opioid-tolerant), opioid-naïve inpatient with new cancer pain and CKD, non-cancer post-operative pain at home (CDC 2022 scope), chronic non-cancer pain on stable opioids (CDC 2022 scope), end-of-life dyspnoea with terminal secretions, refractory terminal delirium (palliative-sedation teaching case). The case form records setting, diagnosis category, opioid status and current regimen (oral MME/day), eGFR and hepatic impairment, goals of care, capacity, surrogate/advance directive, prognosis category and prior adverse effects.
• Doses and rules (sources S100–S120): opioid starting doses only from FDA labels (oral/IV morphine, oral oxycodone, oral/IV hydromorphone; renal/hepatic/elderly wording); breakthrough 10–20% of the 24-h dose and a +50–100% titration suggestion (NCCN); oral:IV morphine 2:1–3:1 and laxatives with opioids (ESMO 2018); conversion ratios 1.5:1 and 5:1 (MASCC-ASCO 2025) with a clinician-entered cross-tolerance reduction (no consensus value exists, so there is no default); CDC 2022 for non-cancer pain only (MME/day is a risk estimate, never a conversion; pause and reassess before ≥ 50 MME/day); naloxone in 0.1 mg increments titrated to respiratory rate (Scottish Palliative Care Guidelines, FDA label); midazolam palliative-sedation steps from the revised EAPC framework (2024). Anything not verifiable (methadone, fentanyl patches, SC opioid doses, second-line sedatives, anticholinergic, antipsychotic and benzodiazepine doses) is labelled "not simulated — specialist decision".
• Run (calm light monitor; 1, 6 or 24 simulated hours per second): pain, dyspnoea, agitation, secretions, nausea, constipation, RASS and respiratory rate (illustrative model output), 24-h trends, a medication administration record with 24-h totals and MME/day, alarms that you acknowledge, and a daily review gate: scheduled and PRN doses run from the regimen you approved, and every titration, rotation or bowel-regimen suggestion waits for your Approve or Decline. HIGH alarms (respiratory rate < 8, unintended RASS ≤ −3) pause the run; nothing is acknowledged for you.
• Palliative sedation (end-of-life module): the controls stay disabled until all six documentation items are complete (refractory symptom with treatments tried and why refractory; goals-of-care conversation with patient or surrogate; consent/assent including depth and duration; interdisciplinary/palliative-specialist input; prognosis — continuous deep sedation only with "hours to days"; jurisdiction/institutional-policy acknowledgement). You choose the target RASS and the symptom, and approve every bolus, repeat and infusion change. The engine never escalates on its own; it raises a "not proportional" alarm when sedation is deeper than the target and a "comfort not achieved" alarm when the symptom persists. Comfort opioids continue on schedule and are never used to sedate. Hydration is a separate, documented decision; vital-sign de-emphasis in comfort-focused care is a labelled teaching toggle.
• Debrief: regimen and totals, per-day opioid totals, review decisions, alarms, the documentation checklist, a proportionality timeline (RASS vs target vs symptom), family-communication notes (shown as plain text, never HTML) and "Download session (JSON)".

Code: data/palliative.js, engine/palliative.js, ui/palliative.js; tests tests/palliative.test.mjs and the palliative block of tests/ui.smoke.mjs; sources in SOURCES.md → "Palliative and end-of-life modules (S100–S120)"; every value in MODEL_CARD.md → "Palliative and end-of-life modules".

Where doses come from

Every dose, range, ceiling, threshold and score cut-off comes from data/drugs.js. Each value traces to an entry in its SOURCES registry: FDA prescribing information, a society guideline or consensus statement, or, for scores and weight formulas, the original paper. The human-readable registry is SOURCES.md. It also records how each source was checked, which drugs were left out and why, and the illustrative model assumptions. Source tags such as [S7] in the UI open the Sources panel at that entry. The auto-filled default within a range is a simulator convention (usually the label's stated usual dose or the lower bound), not a recommendation.

Model limits (honest summary)

• Physiology and PK/PD are illustrative. For most drugs, the compartment constants are fitted so that a default bolus reproduces the onset, peak and duration statements in the label. Where the label gives no timing, the engine uses its own fallback constants. Effect models, synergy, haemodynamic and respiratory gains, patient modifiers and complication magnitudes are engine assumptions. Each drug's actual model record is shown in the Model card (from sim.modelInfo()), so it stays current if a published model is adopted later.
• Inhaled agents use first-order wash-in/out toward the dial setting. Their effect is expressed as a fraction of age-adjusted MAC (Mapleson 1996). Total MAC is the sum of the agents' fractions, which is an uncited teaching assumption.
• The depth number is a processed-EEG depth index (BIS-style, simulated; 0–100, illustrative); it is not a BIS value (BIS is a trademark of its owner). The extubation criteria other than TOFR ≥ 0.9 (Thilen 2023) are illustrative simulator rules.
• No expert anesthesiologist review has been recorded, and no external validation has been performed. The simulator as a whole is not validated.
• The simulator covers adults 18–90 years only. Pregnancy, paediatrics, cardiopulmonary bypass and neuraxial/regional techniques are not modelled.

Relation to the Python Smart NOA controller

With Interlocks (Smart NOA rules) switched on, engine/interlocks.js applies the decision rules of smart_noa_controller.py with the thresholds in config.yaml:

• HR < 48 sets dexmedetomidine to zero.
• MAP < 60 sets dexmedetomidine and lidocaine to zero.
• An unusable sensor value sets the dexmedetomidine, ketamine and lidocaine infusions to zero and turns the status RED.
• Holds latch and are never released automatically.
• The free-text lockouts and the simulation rate ceilings apply (dexmedetomidine 0.5 mcg/kg/h, ketamine 0.2 mg/kg/h, lidocaine 1.5 mg/kg/h). A rate is compared with its ceiling in the Python unit with the same tolerance (1e-9) as smart_noa_controller.py (RATE_CEILING_TOLERANCE), so a rate exactly at the ceiling is accepted in both.

The simulator adds one thing the Python controller does not have: an explicit, logged Resume hold action.

When a label default exceeds a Smart NOA ceiling, the plan editor pre-fills the ceiling and explains why. For example, the dexmedetomidine sedation start rate of 0.6 mcg/kg/h becomes 0.5. With interlocks off, these rules are skipped, but label ceilings and contraindication lockouts always apply. Parity tests (tests/parity.test.mjs) reproduce the outcomes in EXAMPLE_RESULTS.json and the regression cases in test_safety_hardening.py.

Tests and lint

cd simulator
# engine, drug library and Python-parity tests (Node ≥ 18; no dependencies)
npm test

# browser smoke test: starts its own static server (Node http, no dependencies) on a free port.
# Needs Playwright + Chromium (e.g. npm i -D playwright && npx playwright install chromium).
npm run test:ui
#   optional: SIM_URL=http://host:port/simulator/index.html?test=1 (use an existing server), SHOTS_DIR=/path
#   SMOKE_SKIP=1 skips it explicitly when Playwright is not installed; otherwise a missing Playwright FAILS.

# lint (dev dependencies only: eslint 9, @eslint/js 9, globals — `npm install` in simulator/;
# node_modules is git-ignored and is not needed to serve the site)
npm run lint

The page exposes a test hook only when its URL contains ?test=1: window.__smartNoa.snapshot() returns a deep copy (plain data) of the run state; nothing live or mutable is exposed. The smoke test checks:

• the page loads with no console, CSP or external-request errors;
• inline validation works;
• every example case loads and reaches the debrief;
• a GA case runs end to end at 60× (induction → airway → maintenance → emergence → sugammadex → extubation readiness);
• extubating when not ready asks for confirmation and is logged;
• the MAC moderate-sedation case runs;
• guided mode for a GA example (elderly) and the MAC example at 60×: gates appear in order, nothing advances without the gate, Intubate and Extubate are enabled only when their conditions are met, and the debrief notes guided mode;
• "Run full case automatically" for a GA example (MH-susceptible TIVA) and the MAC example at 60×: the banner text, completion, AUTO: log entries, no override, the debrief "ran automatically" note and the session JSON flag;
• a manual control ends automatic mode, but reading a source citation or pre-filling the drug form (drug-row Bolus/Adjust) does not;
• a double-click on a guided-mode gate approves only the gate that was shown (the button is disabled while the approval is processed);
• the extubation dialog pauses the clock while it is open, resumes it on close, and re-checks readiness when a button is pressed;
• technique templates: the selector, the opioid-free plan with the POFA note, lidocaine pretreatment before propofol, and the dexmedetomidine MAC plan with its notes;
• the antibiotic checklist in guided mode waits for the clinician's acknowledgement before Start surgery, and the window alarm is raised at incision;
• the fluids panel (type, boluses, maintenance, totals by type) and the debrief totals (fluids by type, prophylaxis timing, lidocaine including pretreatment);
• no console warnings or errors;
• a sensor fault shows FAULT, turns the status RED/HOLD, and "Resume hold" is logged;
• the palliative modules: the verbatim banner, the ethics panel from every screen, all six example cases end to end (daily review gate holds the clock until every item is decided), the conversion that refuses to proceed without a clinician-entered reduction, the end-of-life sedation gate (disabled until all six items are documented; every titration approved; no automatic escalation; "increase opioid to sedate" refused), free-text notes rendered as text, and every palliative screen at 390 px;
• the Sources, Model card and About panels open;
• there is no page-level horizontal scroll at 390 px.

AI assistance

Code, parameter selection and source gathering for this simulator were produced with AI assistance (Anthropic Claude, model claude-opus-5-5, via Claude Code, November 2026) under the owner's direction. Sources were checked through web-search records of labels and publications as recorded in SOURCES.md; independent human verification of every source by the owner is pending. AI was not used to generate data presented as observed. The owner is responsible for the content.

Files

index.html            page shell (banner, stepper, panels)
css/simulator.css     site palette + dark HMI monitor
ui/app.js             entry: state, stepper, panels, extubation dialog
ui/patient-form.js    step 1        ui/assessment.js   step 2
ui/plan-editor.js     step 3        ui/plan-model.js   plan validation helpers
ui/monitor.js         step 4        ui/charts.js       canvas trends
ui/debrief.js         step 5        ui/model-card.js   Model card panel
ui/sources.js         Sources panel and source tags
engine/, data/        pure simulation engine and drug library (no DOM)
engine/case-actions.js   phase/airway/extubation actions shared by the manual buttons and the sequences
engine/autorun.js        guided mode and the automatic demonstration (pure, testable)
engine/autorun-config.js every guided/automatic condition and timing, with provenance
engine/rescue.js, engine/rescue-config.js  advisory haemodynamic rescue (suggestions only)
engine/depth-advisory.js text-only depth-low advisory (index below the GA band; not an alarm)
ui/advisory.js           advisory cards (haemodynamic rescue; depth index below the GA band)
engine/prophylaxis.js    antibiotic prophylaxis selection, window and redose checks (timing only)
engine/templates.js      technique templates -> plan (doses from recommend())
data/templates.js        technique template definitions (no doses)
tools/model-card-tables.mjs  regenerates the Model card tables (--check verifies they are current)
data/palliative.js       palliative/end-of-life sources (S100–S120), label doses, rules, cases, ethics text
engine/palliative.js     palliative symptom model, dosing/conversion rules, review and sedation gates (pure)
ui/palliative.js         palliative module screens and the ethics panel