Architecture, without the black box
From chart evidence to a reviewable coding decision.
The pipeline deliberately separates generative extraction, bounded judgment, deterministic rules and human control. Each layer can do its job—and no more.
Full data flow
One trace, from source document to signed review.
01De-identified documentImmutable pages and evidence spans
A PDF enters the development workflow only after the operator confirms it is de-identified.
02Atomic clinical factsNo code or modifier selection
OpenAI Structured Outputs turns bounded chart chunks into clinical facts tied to evidence.
03JEV fact validationTyped choices or abstention
The bounded engine classifies facts as performed, planned, historical, ruled out or unsupported.
04Candidate retrievalNo free-form candidate generation
The service date and validated facts retrieve candidates from the active licensed codebook.
05Bounded coding decisionsProbability and evidence retained
JEV selects procedures, diagnoses, modifiers and relationships only among retrieved candidates.
06Coding-line assemblySource and citations preserved
Python assembles supported choices into traceable coding lines and performs unit arithmetic.
07Deterministic enforcementHard constraints cannot be overruled
Active-code, NCCI, MUE, add-on and fee-schedule checks apply as authoritative rules.
08Confidence calibrationRouting signal, not clinical certainty
The weakest contributing decision, abstentions, warnings and hard failures control routing.
09Human reviewAppend-only audit trail
A coder inspects evidence and rules, then accepts or corrects the immutable proposal.
A bounded system can still be wrong