Models are excellent at reading messy clinical evidence, and they stay in the stack. LogicPearl puts a deterministic policy boundary around the final readiness decision.
8 model groups, 24 recorded determinations on identical inputs: 20 approve, 4 deny, with rationales that drift between runs of the same model. Model choice is not a clinical policy. At scale, that variance is rework, leakage, and audit exposure.
| An AI checklist | A LogicPearl artifact | |
|---|---|---|
| Source excerpt per criterion | Often omitted or summarized | Every criterion tied to source PDF text |
| Clause coverage accounting | Usually unmeasured | 99.9% measured on BCBSMA; misses preserved as review work |
| Nested AND/OR & exceptions | Can flatten or invert | Preserved in the decision logic |
| Table semantics | Frequently flattened | ~118k rows interpreted into checkable rules |
| Terminology provenance | Implicit / hidden | Explicit mapping status per criterion |
| Repeatable replay | Can vary on the same packet | Same packet and policy version → same result |
| Reviewer state | Answer only | Promotion gated, sign-off tracked |
| Packet-to-policy evidence link | Not guaranteed | Click-through packet and policy excerpts |
Analysts translating policy PDFs into configuration guess at coverage and take weeks per policy. The Blue Cross Blue Shield of Massachusetts corpus now arrives with coverage measured: 99.9% of clauses, every miss listed, and a semantic diff for every revision. The same accounting can run for any payer policy set.
The status quo is reviewers interpreting policy from memory and audits reconstructed after the fact. The cost is hard to see until an overturn pattern, an audit, or a compliance finding makes it visible all at once.
Watch the pipeline decide a real case, then run it in shadow mode on yours.