REMAND tests a health plan's coverage decision against the measured reality of the law, the precedent, and the regulator — the same test the outside world will apply, run before the letter goes out, not after it's already cost you the dispute cycle.
For plans, TPAs, regulatory counsel, and academic partners. We're inviting a small number of each into the design-partner cohort.
Before the problem exists, we confront the probability of reversal against the measured reality of law, precedent, and regulator — and return the audit before the dispute does. Wrong, we tell you how to correct it. Right, we hand you the argument that defends it. Either way, you're not building the file after the fact.
Class actions against major national carriers allege automated utilization-management systems denied care at scale with minimal individualized review. Whatever their outcome, they've made "how was this denial produced?" a discovery question.
California's SB 1120 now constrains AI in medical-necessity determinations; other states and CMS are moving the same direction. The production process is being regulated, not just the outcome.
An overturned denial costs the plan the care it ultimately pays for — plus the full dispute cycle. Patterned, method-defective denials feed market-conduct exams and bad-faith exposure in the states that recognize it.
We don't publish the detection logic itself — that's the part that has to stay ours. What we do show is the shape of the audit: what each panel checks, and what it hands back.
Every candidate denial is catalogued and classified before it's issued — regime, jurisdiction, and case type routed automatically.
52 criteria mapped to primary-source federal and state law — not "best practice." The actual statute or rule, quoted verbatim where anchored.
The core verdict: issue, correct before issuing, or escalate for human review. Nothing exits the gate without a routed, citable reason.
Up to three real, indexed precedents cited per verdict, drawn from a corpus of 10,603 — never a citation invented to sound plausible.
A calibrated probability with declared provenance. Where the data doesn't support a number, the system says so instead of guessing.
When a denial needs fixing, the system says how to correct it — never how to make the same decision harder to detect. That line is a design rule, not a policy.
A jurisdiction-aware read of bad-faith tort exposure for the governing state — flagged honestly where the underlying case law is still thin.
The defense of a correct denial, written at the moment of the decision — the fact from the record, the law it was tested against, the reasoning — ready before anyone asks for it, not reconstructed months later under pressure.
| State archive | Instrument | Rate |
|---|---|---|
| California (DMHC) — n=21,935 | Built for full opinions that recite the denial ground | 33.4% |
| New York (DFS) — n=21,818 | Built for short summaries that often omit it | 31.2% |
| Washington (OIC) — n=1,528 | Borrowed instrument, declared floor | 37.7% |
Three independently built instruments, converging at roughly one in three. We don't cross-apply one state's ruler to another's archive — a rule enforced in code, not just in the write-up. On the control side: across 55,799 upheld denials with usable text, the same criteria fire in fewer than 1% — evidence the system discriminates, not just detects.
Method-failure measurement complete for California, New York, Michigan and Washington. Oregon supports outcome comparison only — its archive publishes no reasoning text.
Mapped to primary-source federal and state law. The most recent additions are Medicare Advantage (42 CFR 422 Subpart M) and Medicaid managed care (42 CFR 438 Subpart F) — two regimes the engine previously cited without having read, and said so in its own output.
The path to a fully validated release is complete normative coverage across all regimes, then counsel review of the finished set — reviewing criterion by criterion while the base is still growing would mean paying twice for the same read.
No criterion carries a U.S. counsel sign-off on legal application yet. That's the oldest open item on this project, and we say so plainly to anyone we talk to.
49 of 52 criteria only produce a calibrated number against real client outcome data. The first pilot doesn't validate the system — it calibrates it.
Two seats are open right now: a regulatory counsel partner to help close the oldest debt on this project, and a design-partner plan or TPA willing to run the retrospective study on its own book — the first real test of whether detecting a defect can prevent the overturn.