Clinibrium Listen

In a Chilean emergency room, "mareo" is one of the most common complaints and one of the easiest to get wrong. Most dizzy patients have something benign, like positional vertigo. A few are having a posterior circulation stroke, and the clue is often one sentence buried in the history: "ayer vio doble un rato, ahora no." Miss it and the patient goes home with a maneuver for an inner-ear problem. The physician is busy, typing into a form, and the form only knows what was typed. Clinibrium Listen lets the physician just talk. They dictate the case in Spanish the way they would to a colleague. Meta's models transcribe it live, pull out the clinical findings with the exact quote that supports each one, and fill the form while the physician keeps examining the patient. When there is a pause, the system asks the one question that would actually change the decision, like the result of the Dix-Hallpike test. A photo of the patient's medication list adds facts too, like "anticoagulated". The safety call is never made by a language model. A deterministic engine I built before this hackathon (red flags, differential diagnosis, action rails) decides urgency: immediate, priority or outpatient. The models only propose values. Rules in code keep them honest: a finding is accepted only if its quote really appears in what the doctor said, and "sin diplopía" can never become "diplopia present"; unconfirmed proposals can raise urgency but never lower it; a red flag, once heard, sticks until the physician clears it; a kill switch stops every model call and the decision on screen does not change. In the demo's second case, a 68-year-old with hypertension and diabetes and 12 hours of continuous vertigo, the buried double vision becomes an alert, urgency jumps to immediate, the Epley maneuver is blocked, and a clock shows the time since the patient was last known well. A second model acts as a stroke advocate and asks the physician to confirm the finding. At the end the physician gets a referral letter built only from confirmed findings (no model-written prose) and a receipt with a SHA-256 hash they can verify in the browser. I tested it on 12 scripted Chilean cases against a keyword baseline. Muse matched the correct urgency in 12 of 12, the baseline in 11. The one the baseline missed is the case that matters: a wife describing her husband's slurred speech in her own words ("lo encontró hablando enredado"). This is a research prototype, not a medical device, and it uses only scripted dictations. The doctor decides