Notice. Illustration and research prototype. Not clinically validated, not a medical device, not for patient care. Example data are fictional.
Live demonstration. The sliders and test cases work — every score recomputes in your browser. The data is generated, no patient is involved, and nothing here has been verified by a clinician.
Rule governance · clinical panel

Panel review

Every rule in this system is a published threshold applied to a recorded number — KDIGO staging, MEWS, SOFA, the ratio bands, bundle intervals, the nephrotoxin table. No language model. Nothing is generated or guessed. A panel can read every rule and approve, edit, or reject it, which is not possible with a model.

What a clinician sees

Four buttons on every alert. The last two look similar and route to entirely different places — that separation is the whole reason the log is worth keeping.

Sepsis — Assess. Duty physician review within 15 minutes.
Bed 4 · MEWS 4, shock index 1.02, urine 0.38 mL/kg/hr for 7h
rule sepsis_tier_2 v0.1.0 · ISCCM Position Statement, IJCCM 2024

How each rule is performing

Agreement counts "acted" and "later" — both mean the clinician thought the alert was right. "Not this patient" is normal clinical variation; "rule is wrong" is a claim about the system, and only that one accumulates toward review.

RuleStatusResponsesAgreement Not this patientRule is wrong

Raised for panel decision

Rules rejected often enough to warrant review, with the reasons attached. Nothing here has been applied. A closed loop drifts — a unit that systematically under-treats something would teach the system to stop mentioning it, and the drift would be invisible because the alerts simply stop.

Deferral intervals

How long "later" means, per rule. Editable — a deferred urine-output prompt and a deferred sepsis alert have different tolerances.

RuleReturns after