Healthcare decisions should be provable, replayable, and owned by policy, not by whichever model answered that day. LogicPearl exists to make that the default, starting at the decision boundary of prior auth, denials, appeals, and revenue recovery.
The founding observation was an experiment. We gave frontier AI models the exact text of a published payer policy and the same patient packet: the identical inputs, nothing withheld. Each model produced a determination.
Across 8 model groups and 24 recorded determinations, 20 said approve and 4 said deny. Rerun the same model, and the rationale drifted. On one packet and one policy, the answer swung on which model you happened to ask. At scale, that variance is rework, leakage, and audit exposure.
Models are remarkably good at reading messy evidence. The conclusion was narrower and more useful: the decision itself has to come from the policy, deterministically, with a receipt. The model can do everything else.
A public proof trail: live demos, measured policy coverage, and checkable artifacts.
A deterministic decision core. Same input, same artifact version, same answer, with a signed trace you can replay later. It runs in your browser, at the edge, or entirely inside your environment. No hosted SaaS dependency is required.
We compiled 468 Blue Cross Blue Shield of Massachusetts policies from 528 published medical-policy PDFs into checkable criteria and measured the coverage instead of asserting it: 99.9% of requirement clauses, with every miss kept visible as review work. The same accounting can run for any payer policy set.
Working prior-authorization and revenue-recovery workbenches, built against the real workflows of a national payer platform and a revenue-recovery leader. Both run as live demos on this site.
Where we are now: running LogicPearl beside real processes on historical cases, and letting the case-by-case readout make the argument.
Measured on the Blue Cross Blue Shield of Massachusetts policy corpus. The same coverage accounting can run against any payer policy set before a pilot starts.
Healthcare buys on proof, not promises. LogicPearl is engineered so every claim about it can be checked.
The rule, the verbatim source excerpt, the evidence spans, and the artifact hash travel with every result. It can be replayed in an audit two years later, not reconstructed for it.
The engine deploys in your environment: browser, edge, or your own infrastructure. No phoning home is required. Pilots run on de-identified cases or entirely inside your controlled environment, scoped in writing, and you own the artifacts, traces, and readout.
Shadow mode runs LogicPearl beside your current process and delivers a case-by-case agreement report before anything touches production. The readout is the proof.
Run it in shadow mode, review the receipts, and decide from the case-level evidence.
LogicPearl was founded by Ken Erwin, whose background spans both sides of this problem: years of hands-on healthcare work, plus frontier-scale AI training and the HPC clusters behind it at Meta.
That combination is the company's thesis in miniature. When you've trained these models and operated the infrastructure underneath them, you have deep respect for what they do well, and clear limits on what they can promise. LogicPearl exists because the people who know models best are the least willing to let one be the policy of record.
It also means the engineering conversation happens at whatever depth your team needs, from UM workflow down to the runtime your security review will ask about.
Trained frontier models and ran the HPC clusters they train on at Meta, the infrastructure layer most vendors only consume through an API.
Extensive healthcare experience: the policies, the workflows, and the compliance stakes behind every determination on this site.
The demos run in your browser; the pilot runs on your cases.