Distributors designated by Utah’s Workplace of Synthetic Intelligence Coverage should generate and ship a cryptographically signed receipt for each coated AI inference or management analysis as a situation of becoming a member of the state’s AI Studying Laboratory and remaining in it, Glacis Applied sciences stated in an organization announcement printed October 5, 2026. The receipts conform to OVERT (Observable Verification Proof for Runtime Belief), an open technical customary for verifiable proof of AI operations that Glacis created and publishes below a royalty-free patent covenant.
Underneath the settlement, the workplace decides which distributors, methods and occasions are coated, and it writes the requirement into every vendor’s personal settlement with the state; Glacis gives receipt issuance, verification supplies and onboarding, and different proof might accompany a receipt however can’t change one, the corporate stated. The state’s third-party evaluators web page data the Glacis memorandum of understanding as signed August 10, 2026, with a 12-month time period, and the corporate describes the settlement as introduced October 5. Joe Braidwood, the publish’s creator, stated it’s the first time a regulator has written OVERT right into a participation requirement.
What the Receipts Document
A receipt data a management’s reported choice on a coated occasion and the time the choice was made, and it’s signed in the intervening time of the occasion. In line with the announcement, anybody holding the verification supplies can verify that the report has not been altered and that it got here from the anticipated signing key. For example, the announcement describes a scientific assistant configured to route sure requests to a clinician somewhat than reply them itself: the receipt preserves the management’s reported choice on a selected request, giving an evaluator a signed report to look at rather than a abstract written by the seller.
The corporate and the state describe the identical boundary for what a receipt establishes. A verified receipt exhibits {that a} management ran, that it reported a choice, and that the report was not modified afterward; it doesn’t present that the choice was appropriate, {that a} referral reached a clinician, or {that a} affected person obtained acceptable care, and it doesn’t certify security, effectiveness or authorized compliance.
Utah’s evaluators web page units out the state’s model of the mechanism below a class it calls Runtime Attestation: corporations the workplace selects should produce a tamper-evident digital receipt every time a specified security test runs, and the web page states that such a receipt exhibits the test ran with out exhibiting that the AI is protected or that the corporate adopted the regulation. The settlement doesn’t require distributors handy over mannequin weights, uncooked prompts, uncooked outputs or affected person information, as a result of a receipt will be sure to an occasion with a hash or one other identifier that carries no content material.
Utah’s Evaluator Framework and Sandbox Phrases
The Workplace of Synthetic Intelligence Coverage, a part of the Utah Division of Commerce, has signed memoranda of understanding with six unbiased organizations that may evaluation a proposal earlier than it’s authorized and test a pilot’s work whereas it runs. Evaluators report back to the workplace, which retains each choice about whether or not a pilot begins, continues or stops, and the analysis standards are being finalized with every group.
The evaluators web page lists the Coalition for Well being AI, signed October 2, 2026; the Stanford Scientific Excellence Analysis Heart, signed September 30, 2026; Clarion AI Companions and mpathic, every signed September 29, 2026; Vega Well being, signed September 28, 2026; and Glacis Applied sciences, signed August 10, 2026. Every memorandum carries a 12-month time period, and the entries for the Coalition for Well being AI, Stanford CERC, mpathic and Vega Well being state that the time period renews annually.
Underneath the workplace’s printed safeguards, the corporate being evaluated pays the evaluator immediately, the state doesn’t pay for evaluations, and attestation receipts are free to the businesses required to supply them. Evaluators should certify that every analysis was unbiased, report any patient-safety concern to the workplace whether or not or not the corporate passes, and open their work to workplace audits, and a certification that doesn’t match precise findings is grounds for pausing, rolling again or terminating a pilot. No evaluator holds an unique position, and the evaluator agreements grant no regulatory reduction.
The sandbox program’s authorized identify is the Synthetic Intelligence Studying Laboratory, in keeping with the state’s sandbox overview. A participant indicators an settlement with the workplace and with the company that already regulates its subject; the settlement can modify how one particular rule applies throughout a restricted take a look at interval, each different authorized requirement stays in drive, and acceptance will not be an endorsement or approval by the State of Utah. A primary take a look at interval can’t exceed 12 months, and an organization might request as much as two 12-month extensions earlier than the settlement ends and the abnormal guidelines apply in full. Every settlement should spell out the take a look at’s scope, the safeguards the corporate should preserve operating, the particular rule being adjusted, the disclosures owed to customers, and the reporting on which workplace audits run. The workplace can take away a participant instantly for breaking the regulation or the settlement, and it might finish an settlement at any time.
The workplace has additionally signed 5 agreements increasing the laboratory, in keeping with an workplace information merchandise: pilot agreements with Anticipate Health for pelvic flooring bodily remedy, August AI for routine prescription refills and Nolla Well being for zits remedy, alongside grasp agreements with College of Utah Well being and Intermountain Well being that set phrases for approving future AI pilots inside every well being system.
OVERT’s Pending Stewardship Switch
Glacis printed OVERT 1.0 in March 2026 and model 1.1 in June 2026; the specification textual content is free to repeat and adapt below a Artistic Commons Attribution 4.0 license, and the royalty-free patent covenant printed in June stays unchanged.
In a September 21, 2026 publish, the corporate stated that Brenton Hill of the Coalition for Well being AI, Ken Johnston of the AIGovOps Basis and Braidwood signed an OVERT Stewardship and Continuity Settlement on September 14, 2026, below which the coalition and the inspiration would turn into co-stewards holding one vote every and Glacis would turn into a nonvoting editor that can’t alter conformance standards. Glacis said that it competes by itself implementation of OVERT and holds a industrial curiosity in the usual. Within the October 5 announcement, Braidwood wrote {that a} requirement naming one firm’s personal format is weaker than one naming an open customary, as a result of a regulator ought to be capable to confirm a receipt with out relying on the corporate that issued it.
The stewardship switch will not be but in drive, the corporate stated: it takes impact at a time recorded in a separate graduation certificates signed by all three events, and that certificates has not been signed. No launch will be authorized and not using a public remark interval of no less than 30 days and the settlement of each co-stewards, and registry operation is scheduled to maneuver away from Glacis inside 90 days of graduation.
