Sentinel

README excerpt

# Sentinel

**The between-visit clinical review no clinician has time to run — for chronic kidney disease.**

Chronic disease isn't managed in the visit; it's managed in the 364 days *between* visits, where today almost nothing happens until the patient deteriorates and shows up in the ED. Sentinel is an agent that runs that review continuously: it reasons over a patient's longitudinal record, catches a slow decline that's invisible value-by-value, flags a now-indicated treatment the patient isn't on, drafts a referral — and, critically, **knows what *not* to surface**, so a clinician actually trusts it.

CKD is the first wedge: clear labs, clear KDIGO 2024 guidelines, clear deterioration signals, and a real, catchable under-treatment pattern (the non-diabetic albuminuric patient who qualifies for an SGLT2 inhibitor and was never started).

## Architecture

A hard split between deterministic clinical logic and probabilistic language reasoning, human in the loop:

- **Deterministic core** (`core/`) — staging, eGFR slope-fitting, guideline thresholds, suppression rules as tested functions. **Every number in the output originates here.** If a figure appears the core didn't compute, that's a bug.
- **Agent layer** — Claude plans, calls the core functions as tools, sequences the longitudinal reasoning, and composes the ranked action list + referral prose. The model never invents a threshold.
- **Eval harness** (`evals/`) — scores the agent against a 50-patient gold set. Quantified