Cortex Engine does not ask a model to do everything. It combines deterministic clinical checks for facts we can prove with LLM analysis for the parts that require note-level judgment.
Provider-specific JSON shapes are converted into one canonical Cortex contract.
Policy drops unsafe, duplicated, or clinically unsupported model output before it reaches the HUD.
Overlapping findings merge into one card with distinct fix options preserved.
The user sees specific, explainable fixes rather than raw model output or duplicate findings.
Small rules catch facts we can prove from structured chart data. Same input, same output, no model drift.
Models handle contextual judgment: note support, E/M rationale, audit risk, and broader chart review.
Because healthcare QA needs boring guarantees in the places where guarantees are possible. Models are valuable, but they should not be responsible for every obvious rule, every time.
Rules catch the same CKD-stage or demographic mismatch every run.
Cheap checks can stream first while model lanes continue in parallel.
If a model or provider changes behavior, proven checks still hold.
Normalization, policy, and coalescing prevent malformed or unsafe output from becoming UI.
Every layer has a narrow contract: rules, provider adapters, normalization, policy, coalescing, and HUD rendering. When a production failure appears, it becomes a fixture before it becomes tribal knowledge.