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.
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.
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.
| Rule | Status | Responses | Agreement | Not this patient | Rule is wrong |
|---|
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.
How long "later" means, per rule. Editable — a deferred urine-output prompt and a deferred sepsis alert have different tolerances.
| Rule | Returns after |
|---|