Research
Methodology
How CareState approaches AI-assisted denial and appeal intelligence.
Most AI extraction tools evaluate whether a citation is present — whether the system attached a source reference to its answer. That is the wrong bar for denial and appeal review. A system can produce a real, verifiable reference to a payer policy or a clinical note while mischaracterizing what that source actually says. We call this the gap between citation presence and citation correctness.
Citation correctness means the cited span, read on its own terms, supports the specific claim attached to it — including scope (which service, which date of service, which policy criterion) and polarity (a met criterion is not an exclusion). This is the metric that governs how CareState is built and how it should be evaluated.
The CareState architecture separates extraction into three stages: retrieval-grounded extraction that constrains the model to retrieved document spans, a second-pass verification check that examines every claim against its cited text for scope and polarity match, and a human review queue where every finding lands as a draft requiring explicit reviewer approval before it flows into a generated appeal or Case Control File.
The failure mode, made concrete
A synthetic denial. One reason code. Two documents. Two very different answers.
Citation present. Citation incorrect.
“The requested imaging is not medically necessary.”
— cites Payer Policy MED-114
The documented six weeks of failed conservative therapy in the progress note is omitted. The finding is technically cited but materially misstates the patient's clinical picture against the policy's own criteria.
Citation present AND correct. Multi-document synthesis performed.
“The requested imaging meets medical-necessity criteria: the record documents six weeks of failed conservative therapy, satisfying the exception in Payer Policy MED-114.”
— cites Payer Policy MED-114 AND the progress note of record, verified for scope and polarity
Both the policy criterion and the supporting clinical evidence are cited together. The net coverage position is stated correctly. This is the multi-document synthesis case that presence-only evaluation would miss.
The three-stage pipeline
Stage 1
Retrieval-grounded extraction
Documents are chunked at section level using structural pattern recognition (denial reason codes, policy-criteria headers, clinical note sections). Per-category keyword retrieval constrains the extraction model to relevant spans only — it may not generate findings from text it was not given. Every finding must include an exact verbatim quote from the source.
Stage 2
Second-pass verification
Every claim is checked against its cited span for three specific failure modes: scope mismatch (the claim overstates or understates what the span covers), polarity reversal (a met criterion characterized as an exclusion or vice versa), and omission of a material qualifier (a condition, limit, or exception present in the span but missing from the claim). Flagged findings are shown to the reviewer with their flag visible — not hidden.
Stage 3
Human review queue
Every finding lands as a draft requiring explicit reviewer approval. Approval is a recorded human action with reviewer identity and timestamp in an append-only audit log. Generated work product marks unapproved findings as “(pending review)” and carries the footer: AI-assisted draft generated by CareState for professional human review. Human review stays in control.
Output
Generated work product
Appeal Letters, Medical-Necessity Summaries, Payer Question Lists, Denial Trend Reports, Executive Briefs, and Case Control Files — each built from approved findings with source citations. Every document is an AI-assisted draft for professional review.
Read the methodology paper
The full architecture, evaluation protocol, illustrative example, and pre-registered kill criterion are described in the working paper.
Download the methodology paper (PDF)Sammy Orangkhadivi, CareState, Inc. — July 2026.
Early methodology paper. Quantitative results are pending pilot testing on real denial and appeal documents. This paper does not claim peer-reviewed status.