Healthcare AI Evidence Checklist for Section 1557, AB 3030, SB 1120, TRAIGA, and CMS Prior Authorization Audits
Healthcare AI compliance evidence falls into two classes. Runtime evidence is tied to individual AI interactions: policy decisions, disclosures applied, human escalations, PHI handling. Process evidence is documentation: scope determinations, vendor model documentation, reviewer records. Map each regime to both classes, then find which artifacts you cannot yet produce.
Gap assessment checklist
Two classes of evidence
Each of the five regimes asks, in different ways, that an organization show what its AI was permitted to do and what it did. In practice the evidence falls into two classes, and one does not substitute for the other.
Runtime evidence is a record tied to an individual AI interaction or tool call. Examples are the policy decision applied, whether a disclosure was added to an outbound message, whether a request was escalated to a human, and how PHI was handled. Process evidence is documentation the organization or its vendors create outside the interaction: scope determinations, vendor model documentation, policies, reviewer qualifications and pre-deployment reviews.
The mapping in this checklist is an editorial one. It shows where each artifact can originate. It is not a statement that any regulation prescribes a particular record format. Applicability and scope vary by use case, entity type and jurisdiction, so legal counsel should confirm each regime's reach before you build against it.
Section 1557: nondiscrimination and decision support tools
HHS describes its Section 1557 nondiscrimination rule as addressing bias in health technology, covering discrimination on the basis of race, color, national origin, sex, age and disability. The provision at 45 CFR 92.210 concerns patient care decision support tools.
The California Attorney General's February 27, 2026 letter shows how much of this evidence chain sits with vendors. It states that health IT developers must currently provide source attribute information for predictive decision support interventions, including AI-based systems. It also argues that providers' Section 1557 compliance becomes much harder if an ONC proposal removes the model card requirement. That is the Attorney General's stated position, and the ONC change is described as a proposal. Track its status as of November 2026.
For your checklist, model cards and source attribute documentation are process evidence, supplied by developers before deployment. They describe a tool in general. They do not show how it was used in a given interaction. Records of which tool ran, under which policy, with what outcome, are a separate runtime artifact.
CMS prior authorization: decisions, timing and notices
CMS-0057-F requires impacted payers to send prior authorization decisions within 72 hours for expedited requests and seven calendar days for standard requests. The CMS fact sheet excludes QHP issuers on the federally facilitated exchanges from this requirement. The provisions apply to Medicare Advantage organizations and state Medicaid programs, among others, and payers must implement certain provisions generally by January 1, 2026. Payers must also send providers a notice for each decision, including a specific reason for any denial. The rule also ties prior authorization improvements to HL7 FHIR APIs for electronic data exchange.
Three artifacts follow from this. The first is request-received and decision-sent timestamps. The second is the denial notice with its specific reason. The third, where AI or automated tools supported the decision, is a record linking that AI activity to the decision. The first two usually come from the utilization management workflow. The third is where runtime records matter, because an application log rarely ties AI activity to a single case.
Confirm which entity type applies to you before applying these timeframes.
AB 3030, SB 1120 and TRAIGA: disclosure, oversight and scope
AB 3030 addresses disclosure in generative AI patient communications. The core runtime artifact is a per-communication record showing the disclosure was applied. The approved disclosure wording, the communications in scope and any exceptions are process evidence, set by policy and confirmed with counsel.
SB 1120 addresses physician oversight of AI-assisted utilization review decisions. Runtime evidence can show that a request was escalated for human approval and what the outcome was. The reviewing physician's identity, qualifications and clinical rationale, along with the organization's review procedures, are process evidence. A runtime record can show that a human step occurred. It cannot show that the review was clinically sound.
Texas TRAIGA applies according to use case and entity, so scope determination comes first and belongs to counsel. After that, expect a mix: documented purpose and governance decisions from process, and interaction-level records of what the AI did from runtime.
Across all three, PHI handling decisions per interaction (redacted, blocked or allowed) are a runtime artifact. They connect to HIPAA audit controls under 45 CFR 164.312(b). Set retention periods with counsel, because the regimes may not align.
Where runtime policy enforcement fits
Runtime enforcement addresses only the first class of evidence. It does not replace vendor documentation, scope analysis or physician review procedures, and it does not by itself establish compliance with any of these regimes. Evaluate it on whether records attach to individual interactions, whether retention and data residency are configurable, and whether the enforcement point adds acceptable operational dependency to your AI path.
Trussed enforces policy in the execution path: allow, redact, block, or escalate for human approval, for individual interactions and tool calls. That enforcement generates audit-ready evidence mapped to HIPAA. Trussed runs in production with healthcare organizations and can be self-managed in the customer's cloud, with configurable retention and data residency. It does not provide model provenance or end-to-end tracing of agentic workflows, and it does not promise regulatory outcomes.
A practical next step is to take one AI-supported utilization review or patient communication workflow and trace which of the six artifacts above you can produce today.
Evidence origin by regime
Section 1557
Vendor model documentation (process) plus per-interaction records of tool use (runtime)
AB 3030
Disclosure applied to each generative AI patient communication (runtime)
SB 1120
Escalation and approval events (runtime) plus physician review records (process)
TRAIGA
Scope confirmed by counsel (process) plus interaction-level records (runtime)
CMS prior authorization
Decision timestamps and denial-reason notices tied to AI-supported decisions (mixed)
Review your evidence gaps against runtime enforcement
See how policy enforced in the execution path produces audit-ready evidence for individual healthcare AI interactions and tool calls.
See How Evidence Is Produced