Why triage rules belong in code, not prompts
Burying safety logic in a system prompt is a liability. Here's how CareGuard makes every urgency decision auditable.
By CareGuard Team
Large language models are remarkable at understanding what a patient is describing. They are far less reliable at consistently acting on it the same way every time. That gap is exactly where patient safety lives — and it's why CareGuard keeps the triage decision out of the model entirely.
The problem with prompt-based safety
A prompt that says 'always escalate chest pain to emergency care' works most of the time. But 'most of the time' is not a standard you can defend to a clinical governance board. Prompts can be overridden by clever phrasing, drift across model versions, and leave no record of why a decision was made.
What CareGuard does instead
The model has exactly one job in triage: extract the patient's symptoms faithfully, using their own words. Everything after that is deterministic code.
- An explicit red-flag list scans the extracted symptoms.
- On a match, urgency is forced to EMERGENCY — the model cannot override it.
- When the picture is ambiguous, scoring escalates up, never down.
- Every decision, and which red flags fired, is logged for audit.
The model proposes; the rules decide. That separation is the whole product.
The result is a triage layer you can actually reason about, test, and prove — not a black box you hope behaves.