Hammer Labs tests when healthcare agents are wrong, when they should refuse, and where human review still belongs.
Pick a billing code and an insurer. This answer was served from public payer evidence on SEP 6, 2026, copied unedited. Missing evidence returns no figure; a published rate never becomes a paid amount or a coverage decision by implication.
Hammer Labsrate_position with the same code against hammer-coverage-intelligence:2026-08-31+11.rate_position(code: "91200", payer: "aetna") N COUNTS · CURATED TREE One row per distinct rate, layered re-pulls removed. n is a count of contract rows in this extract, not a count of transactions or of providers who perform the priced service, and it has not passed a claims-joined ghost-rate or plausibility filter.
No cell was served for this payer and code in this extract. Payers omit, fragment and misfile disclosures market-wide, so the supportable sentence is that no rate was recorded in this month’s disclosure, not that the code is not covered.
Medicare Physician Fee Schedule, vintage 2026-08-21: 1.01 RVU.
Priced on the Physician Fee Schedule as relative value units, not dollars. Converting relative value units to a dollar figure needs the annual conversion factor and a locality-specific geographic adjustment, neither of which this world holds, so no ratio to the anchor is computed for this code.
Most healthcare AI still ends with a nurse, coder or medical director checking every output. Real claims rarely carry a clean answer key. Hammer builds one from preserved evidence, then measures where an agent is wrong and where it should have refused. Payers now face decision deadlines and audit requirements under CMS-0057-F; review must become measurable, not disappear into another queue.
Capture each source with its date, basis and original artifact before upstream files are replaced.
Run agents against records where expected decisions and failure conditions exist before any model responds.
Return unknown when evidence is missing. Name what would settle the question instead of laundering absence into confidence.
115 billing codes, 9 payers, Medicare fee schedules, coding edits and coverage articles. Install it, run a query and inspect its receipt. Public material demonstrates behavior; it is not the paid product. The CMS-derived archive behind it is free to read in the browser, dated vintages included.
Same tests, dated record and refusal boundaries on your evidence. Hosted by us or inside your VPC. Our team works alongside yours until your people run it. Our hours on your account go down every quarter. We show you the number.
Bring one code family or one cohort. We test where the agent is wrong, where it should refuse, and what measuring the same behavior on your claims requires.
Useful work becomes a private world on your evidence, priced annually. Bespoke findings become reusable tests, configurations and refusal classes. Work that cannot cross that boundary stays outside the platform rather than becoming permanent custom code.
Assessment fee comes off year one. Your team can run the system on our infrastructure or inside your VPC.
Each names the unsupported sentence and, where known, evidence that would settle it.
Claims, payer and provider data, clinical infrastructure, lending, insurance and defence. Not domain expertise added after the product.
Every release and preserved vintage gets a dated entry. Nothing planned appears as shipped.