Decision vs Generative Language Models for Medicare Prior-Authorization Documentation Review

Prior-authorization review for Medicare outpatient services includes a documentation step in which the clinical packet is checked against a policy checklist and each finding must be supported by a quotation from the record. Generative language models can perform this step well but write their own quotations, which may not match the source. This report compares a generative model (Gemini 2.5 Pro inside Palantir Foundry) with a decision model (TypeSafe's Jev), which answers fixed multiple-choice questions with probabilities, so that a cited sentence is verbatim by design. Both run inside the same routing and evaluation pipeline on 39 synthetic requests with designed outcomes. The engines reach the same documentation headline on 39 of 39 requests and agree on 93% of 256 checklist items. The generative model failed to produce a verbatim quotation on 8 items; the decision model failed on none and reported a probability on every verdict. The decision model's median latency per request was about 10 times lower, 8 seconds against 81. All requests, patients and notes are synthetic; no real patient data was used. The benchmark is exploratory in size and does not establish performance on real clinical records or other document-review tasks. Technical report by AKOS AI.