Third-Party Evaluators

The Importance of Independent Oversight

Companies testing AI in Utah's AI Sandbox are required to report on their own performance. However, our office does not rely on those reports alone. We have signed agreements with independent organizations that can review a proposal before it is approved and check a pilot's work while it runs, so that there is shared trust and accountability.

Approved evaluators report to our office, and we make every decision about whether a pilot starts, continues, or stops. The evaluation criteria are being finalized with each organization now. For the pilots themselves, see Authorized AI Pilots.

An Emphasis on Post-Deployment Evidence

Utah's AI Regulatory Sandbox is built on a fundamental reality: static pre-deployment testing alone cannot guarantee how complex AI systems will perform in real-world clinical environments. Even if a model passes pre-deployment testing with flying colors, AI models, patient populations, and user behaviors naturally evolve over time. Our oversight framework places the greatest weight on continuous, post-deployment data collection and real-world quality assurance.

Pre-deployment evidence, including synthetic data, historical back-testing, simulated vignettes, and shadow-mode testing, remains essential for demonstrating baseline safety. The more robust a participant's pre-deployment evidence, the shorter their Phase 1 calibration period will be. However, because real-world deployment surfaces unique edge cases and behavioral dynamics, ongoing post-deployment monitoring ensures that safety and clinical efficacy are continuously verified as these systems operate in practice.

What Third-Party Evaluation Looks Like

Scroll sideways to see all four stages.

How the Third-Party Evaluator Program works: what the applicant, the third-party evaluator, and this office do at each of four stages
1. Apply Applicants submit pilot application with evaluation plan and expert review(s). 2. State Review OAIP reviews application, evaluator materials, and overall risk and readiness. 3. Pilot Deployment Approved pilot deploys in the real world, with third-party evaluation underway. 4. Ongoing Governance Evaluator delivers evidence to the state. OAIP decides whether the pilot continues, is modified, or advances.
Applicant
  • Create a pilot deployment plan
  • Propose outcome metrics to share with the state
  • Contract with a state-approved evaluator (from existing pool, applicant-suggested partner, or a mix)
  • Respond to questions
  • Refine plan as needed
  • Implement the pilot plan
  • Serve real Utahns
  • Generate performance data according to the plan
  • Applicant continues to operate responsibly
Third-Party Evaluator
  • Provide expert review(s) assessing conformance with OAIP standards
  • Be available to answer state questions as needed
  • Conduct independent evaluation, primarily by sampling post-deployment interaction data
  • Provide findings to the state in accordance with OAIP standards
  • Periodically evaluate participant claims as needed per the deployment plan
OAIP
  • Maintain evaluator approval pool (no privileged certifiers)
  • Set evaluation and evidence standards
  • Decide whether to grant the pilot
  • Retain full control over which pilots are approved, and under what terms
  • Monitor pilot progress
  • Receive and review evaluation data
  • Determine whether the pilot continues, is modified, or advances to the next phase
  • Continue oversight
  • An open evaluator network

    The program uses a market-based pool of qualified third parties. Evaluators compete on an equal basis.

  • Expands capacity

    Leverages external expertise without growing government.

  • Focused on real-world evidence

    Primary goal is an audit. Objective confirmation of key participant claims, primarily from post-deployment data.

  • Enables broader impact

    Evidence can help support interstate and federal oversight, supporting efficiency and shared standards development.

For more details, see our Evidentiary Expectations Page.

Three Kinds of Evaluators

  • Clinical and technical evaluation

    Before a pilot is approved, an evaluator can test the AI for clinical errors and unsafe behavior, have clinical experts review the proposal, and check privacy safeguards. Once a pilot is running, it can audit samples of the AI's real activity and check the company's reports against them.

  • Evaluator network

    The Coalition for Health AI does not evaluate pilots itself. It qualifies evaluators from its partner program to do this work, under the same independence rules as every other evaluator.

  • Runtime attestation

    Companies this office selects must produce a tamper evident digital receipt each time a specified safety check runs. A receipt shows the check ran. It does not show that the AI is safe or that the company followed the law.

Safeguards in evaluator agreements

  • The company being evaluated pays the evaluator directly. The state does not pay for evaluations, and attestation receipts are free to the companies required to produce them.
  • Evaluators must certify that each evaluation was independent, report any patient safety concern to this office whether or not the company passes, and allow this office to audit their work.
  • A certification that does not match the evaluator's actual findings is grounds to lose recognition.
  • No evaluator has an exclusive role. A company directed to get an evaluation can choose any evaluator this office recognizes.

Approved Third-Party Evaluators

Each agreement is a memorandum of understanding with this office.

Coalition for Health AI

Evaluator network

A national coalition that develops shared standards for health AI. The Coalition for Health AI does not evaluate pilots itself. Under this agreement it qualifies evaluators from its partner program to work in the Learning Laboratory under the same independence rules as every other evaluator.

Signed
October 2, 2026
Term
12 months, renewing each year

Stanford Clinical Excellence Research Center

Clinical and technical evaluation

A research center in Stanford University's Department of Medicine that designs and tests new ways of delivering health care, including clinical uses of AI.

Signed
September 30, 2026
Term
12 months, renewing each year

Clarion AI Partners

Runtime attestation

A Salt Lake City firm of lawyers and data scientists that tests and audits AI systems. Under this agreement it issues signed receipts showing that specified safety checks ran.

Signed
September 29, 2026
Term
12 months

mpathic

Clinical and technical evaluation

Empathy Rocks, Inc., doing business as mpathic. Recognized to evaluate AI systems before and during a pilot, including clinical review of AI conversations.

Signed
September 29, 2026
Term
12 months, renewing each year

Vega Health

Clinical and technical evaluation

Vega Health Inc. Recognized to evaluate AI systems before and during a pilot, including testing for clinical errors and auditing live AI activity.

Signed
September 28, 2026
Term
12 months, renewing each year

Glacis Technologies

Runtime attestation

A Seattle company that builds tamper proof records of AI behavior. Under this agreement it issues signed receipts, built on the openly published OVERT standard, showing that specified safety checks ran.

Signed
August 10, 2026
Term
12 months

Frequently Asked Questions

What's the purpose of Third-Party Evaluators?

Third-party evaluation is how this office gets independent evidence about AI pilots in the Learning Laboratory. Qualified evaluators review a company's plan before a pilot is approved, then check the company's key claims against real-world data once the pilot is running. Our office uses that evidence to decide whether a pilot continues, is modified, or advances.

Who decides whether a pilot is approved?

The Office of Artificial Intelligence Policy (OAIP), in coordination with state regulating agencies like the Division of Professional Licensing (DOPL), retains 100% of final decision-making control. While third-party evaluators provide independent technical verification and clinical audit reports, their certifications do not replace state regulatory judgment or guarantee sandbox admission, phase advancement, or approval.

Are third-party evaluators involved pre-deployment, post-deployment, or both?

Third-party evaluators are involved in both pre-deployment vetting and post-deployment oversight.

  • Pre-Deployment (Vetting Audits): Evaluators conduct clinical SME reviews, red-teaming (testing for clinical errors and guardrail bypasses), HIPAA compliance assessments, governance structure reviews, and pre-launch safety benchmark validation.
  • Post-Deployment (Oversight Audits): Once a pilot is live, evaluators audit de-identified conversation logs, run random sampling on real-world AI decisions, monitor for model/population drift, and verify participant reports against live activity logs.

What does an evaluator actually check?

It depends on the kind of agreement:

  • Clinical and technical evaluation: before approval, testing the AI for clinical errors and unsafe behavior, expert review of the proposal, and privacy checks. During the pilot, auditing samples of the AI's real activity and checking the company's reports against them.
  • Evaluator network: qualifying partner evaluators to do that work.
  • Runtime attestation: issuing a tamper evident receipt each time a specified safety check runs.

How does a company choose an evaluator?

A company applying for a pilot contracts with a state-approved evaluator. It can pick from this office's existing pool, suggest a partner of its own for approval, or use a mix. No evaluator has an exclusive role: a company directed to get an evaluation can choose any evaluator this office recognizes.

Do evaluating vendors have exclusive contracts with the State or sandbox applicants?

No. All third-party evaluator agreements executed by the Office of Artificial Intelligence Policy (OAIP) are strictly non-exclusive. No evaluator holds an exclusive right to assess sandbox pilots, and sandbox applicants directed to undergo third-party review are free to select any evaluator recognized by the Office.

Is the list of approved third-party evaluators fixed?

No, the evaluator network is open and expanding. The OAIP maintains full authority to recognize additional independent evaluation organizations, academic clinical centers, and technical auditing firms as long as they satisfy the state's minimum standards for independence, technical capability, and clinical rigor.

Who pays for the evaluation?

The company being evaluated pays the evaluator directly. The state does not pay for evaluations. To protect independence, evaluators must certify that each evaluation was independent, report any patient safety concern to this office whether or not the company passes, and allow this office to audit their work. A certification that does not match the evaluator's actual findings is grounds to lose recognition.

Do companies have to expose their intellectual property (IP) or proprietary source code to evaluators?

No. Evaluation engagements do not require applicants to disclose model weights, proprietary source code algorithms, or underlying trade secrets to evaluators. Third-party evaluation methodologies rely on privacy-preserving techniques, such as testing against de-identified data sets, synthetic data, in-silico simulation encounters, and output/decision log analysis. Furthermore, sandbox participants retain full ownership of all pre-existing intellectual property under state regulatory agreements.