It does not arrive as clean fields.
It arrives as faxes, scans, copied-forward chart language, and handwriting. Before anyone can apply the policy, the case has to become evidence a reviewer can inspect—not just a summary a model produced.
The job: preserve where every fact came from while turning the packet into reviewable evidence.
Prior auth · lumbar radiofrequency ablation
24 recorded determinations
Three recorded runs per model group. Haiku changed its own answer; matching verdicts still cited different criteria.
The model is useful precisely where it is not final.
OCR and clinical language models find candidate evidence in the packet. But a model's answer can drift across vendors and reruns. LogicPearl keeps that probabilistic work in the evidence step, where every finding can be verified.
The boundary: models may propose facts; they do not become the policy of record.
Five questions replace one opaque answer.
The compiled policy artifact evaluates the same criteria the reviewer has to defend: what is satisfied, what is missing, what conflicts, what blocks the case, and what would make it ready.
What would make this ready: an MRI dated within 90 days of the request, and reviewer sign-off on the laterality conflict.
Request missing evidence
Records request drafted for imaging; laterality conflict routed to the nurse-review queue.
§2.b · imaging recency · p.4 of source PDF
quote hash sha256:9f3c...a41e
The answer never asks for blind trust.
The reviewer sees the case and policy together. Each criterion links back to the packet location and the verbatim source-policy excerpt, including OCR QA state for uncertain or handwritten evidence.
- Evidence boxes anchored directly to the scanned page
- Policy requirements linked to source text and page
- Agree, flag, and reject states for every proposed match
The reason you gave is the reason you can defend later.
The packet, evidence spans, rule, source policy, reviewer history, and exact artifact version stay bound together. If a policy changes, the semantic diff explains why the same case would route differently—without rewriting history.
Why an answer changed is a question you can answer.
Run the same case against the April source set and it routes to review. No covering policy existed yet. Run it against June, where a new policy took effect on the first, and it decides. Same case facts; the source set changed. That is an explanation an auditor accepts, and a model can't give.
When a payer updates a policy PDF, the change lands as a semantic diff between artifact versions, not a silent behavior shift. You see which open cases just changed, before they surprise you.
"You can't defend a denial with a reason you never gave."
Move the goalposts mid-appeal and the process starts over: new disclosure, new clock, new exposure. And the stakes compound: industry-wide, most appealed denials are overturned, and overturn rates feed CMS Star Ratings, which drive bonus payments and enrollment. LogicPearl keeps the original rule, evidence, and source bound to the decision, so the reason you gave is the reason you defend.