Artificial Intelligence
for clinical reasoning research.

A care team, not a chatbot.

Real clinical encounters are not a single back-and-forth with one person. A trainee — alone or working as part of a team — moves between the patient, the patient's family when the patient cannot give a history themselves, the bedside nurse who holds the medication record and the recent vitals, and the on-call consultant brought in for a second opinion. Each interlocutor is a distinct AI agent with its own persona, voice, knowledge scope, and arrival cue. Conversations branch and reconverge in a single session transcript, with the dialogue fidelity an actual clinical interview demands.

• Patient interview — natural-language history-taking from a model that holds the case's clinical narrative, stays in role, and only volunteers what a patient at this stage of the encounter would plausibly volunteer.
• Family or surrogate — for the patient who cannot speak for themselves (unconscious, intoxicated, paediatric, post-arrest, dementia, aphasia), the family agent carries the collateral history. They know the home medications and what happened on the way to the ED; they do not know the labs.
• Bedside nurse and on-call consultant — the nurse holds the running medication record, the most recent set of vitals, and what was tried in the last few minutes. The consultant is paged in with a 1–3 minute server-anchored arrival and brings the broader differential. Per-case rosters script who is present from the start and who arrives on what cue.
• Solo or team problem-solving — a single trainee orchestrates the whole team in one session; cohort or team training reviews and analyses the same transcript exactly as the dialogue happened. Sessions are recorded turn-by-turn for replay and debrief.
• Knowledge scope is admin-controlled per agent, per case — the patient does not know the consultant's notes, the family does not know the labs, the nurse does not hold the post-discharge plan. The information asymmetry is the realism.
• Local LLM backends (LM Studio, Ollama) are first-class; cloud providers are optional. Runtime model switching with per-user usage tracking.

Designed to deteriorate.

Cases run on a time-keyframed scenario engine with explicit down-going trajectories. Vitals drift toward decompensation unless the trainee intervenes appropriately. Alarms cross severity thresholds, banners surface, audio escalates; acknowledgement and snooze route the alarm through audit history rather than silencing it.

• Time-keyframed vital trajectories with interpolation between frames; the engine ticks until trainee action redirects the course.
• Severity-bounded alarms — urgent, beep, chime, silent — routed simultaneously to banner, toast, audio, and audit surfaces.
• A central notification dispatcher with one routing matrix per role; no parallel toast/banner/alarm paths to silently disagree.
• Cross-case acknowledgement clearing — case A's acknowledged alarms never silence brand-new alarms in case B.

Every event row carries the patient's vitals — and the trainee's affect.

The session is not a black box. Every navigation, examination, order, treatment, message, and alarm response is captured as a verb-typed event; each row is enriched server-side with the active room, the contemporary vital signs — heart rate, blood pressure, SpO₂, respiratory rate, temperature, EtCO₂ — and the corresponding affect window at the moment the action occurred.

• Verb-coded action stream merged into clinical labels for downstream analysis.
• Each row stamped with the active room (Patient · Examination · Laboratory · Radiology · Consultant) so behavioural trajectories segregate by setting.
• Vital-sign columns on every event let behavioural sequences correlate directly with physiologic state — no offline merge required.
• Affect columns on every event row (valence, arousal, dominant emotion) close the loop on multimodal analysis.
• Cohort export in CSV or JSON: login logs, chat logs, complete-session bundles, questionnaire responses.

Install in the shape that fits the site.

Four install paths and two upgrade paths cover the realistic deployment surface — from a single classroom laptop to an air-gapped institutional host. The same operator tool handles in-place upgrades with snapshot and automatic rollback on failure.

Full operator manual: docs/INSTALL.md · docs/DEPLOY.md · docs/UPDATING.md · migrations/MANIFEST.md