November 11, 2026
The Honorable Gavin Newsom
Governor of California
1021 O Street, Suite 9000
Sacramento, CA 95814
cc: Secretary Nick Maduros, California Government Operations Agency; Director Caroline Thomas Jacobs, Governor’s Office of Emergency Services; Assembly Speaker Robert Rivas; Senate President Pro Tempore Monique Limón; Assembly Privacy and Consumer Protection Committee Chair Rebecca Bauer-Kahan; Senate Privacy, Digital Technologies, and Consumer Protection Committee Chair Christopher Cabaldon; Sen. Jerry McNerney
Dear Governor Newsom:
The Partnership for Assurance, Credibility, and Trust on AI (PACT AI) thanks you for your leadership in enacting a framework for independent AI oversight in California and for inviting substantive input to inform its swift implementation. Artificial intelligence is redefining how the American economy works: how healthcare gets delivered, how retailers serve customers, how businesses run. But as frontier models advance and new capabilities emerge, our ability to verify that these systems work as designed has not kept pace.
PACT AI was created to close this gap and ensure that trust in AI scales alongside innovation. PACT AI unites dozens of independent verification organizations (IVOs), companies deploying AI, and civil society leaders to advance the science of AI evaluation and build the infrastructure necessary to ensure AI can safely and responsibly deliver on its potential.
Companies and consumers can’t be expected to take it on faith that AI models and systems are working as intended. But, if they can verify that AI performs safely, securely, and reliably, the economy will adopt the technology with confidence and at scale.
We’ve seen this pattern in other high-stakes industries. Financial markets run on independent audit. Automobiles, airplanes, and electrical products run on independent safety testing. In each case, a professional assurance sector gives buyers, insurers and the public a basis for trust that does not depend on the seller’s word. That trust is what allowed those industries to grow and allowed our society to reap the benefits of the technology.
PACT AI’s members are already leading the way, doing the hard work of evaluating AI systems and building out the robust AI assurance ecosystem our economy needs. Your Executive Order N-9-26 presents an opportunity for regulation to catch up to where the market already is and legitimize this vital work.
PACT AI is encouraged to see California champion an AI assurance model that serves as a blueprint for the rest of the nation. SB 813 (McNerney; 2026) and AB 1405 (Bauer-Kahan; 2026) establish much of the initial architecture with a keen focus on independence and competence. But for independent AI oversight to be effective, it must be implemented thoughtfully. Through this Executive Order, the state has opportunities to clarify the requirements to establish genuine independent AI assurance.
Based on the extensive expertise of our members, PACT AI has attached a number of recommendations to ensure that the IVO and audit ecosystem for AI earns trust with consumers and companies alike, while enabling continued responsible innovation here in California. These include recommendations for (1) securing the four conditions for credible, independent AI evaluations and audits; (2) establishing a robust IVO ecosystem for frontier AI; and (3) growing a competitive marketplace for innovation.
We appreciate your consideration and look forward to continuing to engage with state leadership on AI assurance.
Respectfully,
Bri Treece
Executive Director
Partnership for Assurance, Credibility, and Trust on AI (PACT AI)
PACT AI’s recommendations for implementing the IVO framework
While SB 813 and AB 1405 establish much of the initial architecture for AI assurance, the Executive Order gives the state the opportunity to flesh out key details to ensure these laws are implemented effectively, as well as explore ideas for furthering the scope.
Given how fast the technology is evolving, it’s essential that the state’s assurance mechanisms remain flexible enough to keep pace. We have much of the scaffolding to get started, and, as is the case with every assurance ecosystem that’s been built for high-risk industries, early implementation will need to evolve as the system matures.
It will be important to take an iterative approach and evolve the system to meet the need for responsible innovation and adoption of AI here in California.
As PACT AI has engaged with state leaders and the expert panel, we’ve focused on providing tangible, actionable implementation recommendations based on how the IVO ecosystem is working today, how it can mature to meet the unique needs of frontier AI, and how effective assurance systems have worked historically in other contexts. While some of our recommendations may apply to the broader AI IVO and audit ecosystem, this document is tailored to address the questions in the Executive Order regarding frontier AI.
What follows is a summary of our contributions to the panel’s report.
Securing the four conditions for credible, independent AI evaluations and audits
We recommend that the state ensure evaluations and audits conducted by IVOs meet four core conditions: independence, appropriate expertise, sufficient evidence and access, and rigorous, state-of-the-art methodology.
SB 813 and AB 1405 already lay the groundwork for independence and competence, but there is an important opportunity to establish and codify clearer expectations around access and evidence, as well as the degree of independent testing required to turn the review of developer evaluations into genuine independent assurance.
Independence.
Robust assurance requires the IVO to exercise independent judgment over planning, execution, evidence, and findings, rather than simply validating developer conclusions or serving as a rubber stamp. This includes financial independence: compensation must not depend on the findings or outcome of the evaluation, and financial relationships must not influence evaluator judgment. Implementation should ensure independence extends in practice, not only in formal organizational relationships, consistent with principle 2 of AEF-1, the AI Evaluator Forum’s standard for independent third-party evaluations.
Expertise appropriate to the risks being assessed.
AI evaluations and audits may require different expertise depending on the risk domain and on whether the subject is model capabilities, safeguards, or governance. The standard should be engagement-specific competence (such as engagement areas outlined here and here) and domain-specific expertise, not merely generic technical qualification (§8898.1(c)(2)(A)–(B); §11549.83(f)(1)(F)). California could require IVOs to document how each engagement team has the necessary mix of frontier-model, risk-domain, and assurance layer expertise.
Sufficient and appropriate evidence through adequate access.
Strong assurance is impossible if IVOs see only selected summaries or developer-curated results. The IVO should determine what information and system access is reasonably necessary to support its conclusions. Depending on scope, that includes underlying evaluation results, code and configurations, black-, white- or gray-box model access, safeguard tests, monitoring and incident data, system logs, and relevant personnel. AB 1405 requires reports to disclose material gaps in evidence, systems or access (§11549.83(d)(1)(E)), but does not itself guarantee sufficient access. California should clarify that IVOs must obtain the evidence reasonably necessary to support their conclusions, and that material access limitations or insufficient evidence should lead to a qualified or limited conclusion rather than being worked around. See Section 4.2 of this paper for additional recommendations in this area. Importantly, some evaluations are only effective with extended-duration physical access. Two examples are i) evaluations of the AI model development process (versus just evaluation of the completed product) and ii) evaluation of progress towards Recursive Self-Improvement (RSI). In both cases observation of decision-making processes will be critical to effective oversight.
State-of-the-art methodology and meaningful independent verification.
Evaluations should be grounded in an explicit threat model, use methods appropriate to the capability, safeguard or control being assessed, account for known limitations, document uncertainty, and evolve with AI evaluation and auditing science. For material claims, IVOs should perform independent verification, such as inspecting underlying evidence, rerunning selected evaluations, stress-testing assumptions or conducting independent tests, rather than relying solely on developer representations. The main gap in both statutes is that neither specifies how much independent testing is required. California should require methods to reflect the latest science, while avoiding fixed benchmark mandates that may quickly become obsolete.
Establishing a robust IVO ecosystem for Frontier AI
Get started now and leave room for iteration
Waiting until the frontier AI IVO ecosystem is fully “ready” is a chicken-and-egg problem: the market won’t grow as quickly as it can without clearer demand. There are multiple organizations doing independent evaluations and audits today and many more starting or pivoting to focus on frontier AI. We have enough to get started and the sooner we genuinely engage this system, the faster it will grow.
Here’s what the state can do now to move this ecosystem forward:
Mandates create the market.
Voluntary approaches haven’t produced enough demand to support meaningful growth in the IVO market. As regulatory requirements bring in IVOs and push evaluations, assessments, and audits to become ongoing processes rather than one-off tests, more IVOs will form and the market will grow to meet the demand.
Sunrise requirements.
The state could choose to set a fixed (near) future date for the mandate so that developers and IVOs have time to prepare and to incentivize early adoption. The state might consider an extension if there are not enough designated IVOs or enough capacity amongst them to meet the deadline. This can also be mitigated with some approaches below, such as allowing for scope limitations and professional judgment in safety cases.
Stagger mandates by risk area.
The state could choose to prioritize or pull forward certain mandates based on both topic urgency and IVO readiness or capacity. Mature issue areas with more mature evaluation and auditing capabilities (cybersecurity, privacy, governance and process) can be required now. The highest-risk areas (RSI, loss of control, and control verification) should be prioritized because they are urgent, even though supply is thin. A mandatory requirement incentivizes new IVOs to enter the market and provides new and existing IVOs with certainty of demand so they can invest more in capacity-building.
Allow for scope limitations.
Under assurance standards such as ISAE 3000, assessors can report criteria they couldn’t verify. This keeps the requirement mandatory while being honest about current limits. Framework-adherence verification and general model evaluations can follow as standards mature.
Leverage professional judgment and safety cases.
Detailed, prescriptive standards are not always a prerequisite for credible safety inspection. U.S. nuclear safety oversight has long relied on professional judgment within defined requirements and review procedures, while offshore safety regulation uses performance-based requirements assessed by accredited independent auditors (GAO, BSEE). These precedents provide a practical model for requiring frontier AI laboratories to produce safety cases (i.e. structured arguments supported by evidence) that competent, independent inspection bodies can scrutinize for completeness, soundness and adequate risk control. OpenAI’s proposed approach supports this direction, explicitly recommending that auditors receive sufficient access to examine safety-case claims and identify gaps. A safety case-based approach could provide a bridge between informal, ad hoc methods and formal standards.
These are just a few options for implementation that balance the need for action now with the practical realities of an evolving technology and evaluation science. Any implementation approach will need to be flexible and have built-in mechanisms to adjust course.
Even while the standards and procedures are being developed, it’s important to remember that independent presence has value in itself. Having an outside party observing impacts behavior and increases accountability, even before full standards exist.
Growing a competitive marketplace for innovation
The current risk to frontier AI IVOs’ independence is structural. Since there are currently only four to six frontier developers, an IVO focused primarily on frontier work will likely get more than 20% of its revenue from a single developer. Revenue concentration is a recognized area of risk in the financial audit profession, and this level of reliance on a single audit client would be considered an impairment to independence. The main concern is that auditors would feel pressured to “please” the audit client to avoid losing future business.
To address the threat of retaliation influencing IVO work, either of these must happen:
The share of revenue associated with any specific frontier developer must be reduced.
The threat of retaliation must be reduced. Others have researched these same issues in the past, and the remedies have been explored in several other industries.
Lowering revenue concentration would require the number of frontier developers to significantly increase, or IVOs diversified enough to earn significant revenue from other types of organizations. Both could happen as the AI and IVO markets mature, so the state should keep its processes flexible.
A concern on the other side is that mechanisms that limit revenue concentration or retaliation could create an overly bureaucratic system that slows the growth of a competitive IVO marketplace. That marketplace will be a major part of the AI industry and will need continual innovation, which in turn requires funding, talent, and incentives for “big bets.” We therefore encourage the state to start with a process that allows learning and adapting as the market, science, and technology evolve.
To reduce or eliminate the threat of retaliation, the following approaches could be effective:
Remove IVO selection authority from the developer, but allow the developer to have a voice in the process.
A separate body would decide which IVO assesses which risks for each developer, minimizing the potential for retaliation. The body could be the state, or it could be an independent third-party administrator designated by the state for this role. If the state does have selection authority, we recommend that the process include input from the developers who have unique understandings of the risks and their organizations, and an outside oversight body should review decisions periodically. There is a risk that this approach becomes slow-moving, bureaucratic and a disincentive for new entrants to the IVO marketplace. We encourage the state to make the process as streamlined, flexible, and easy to navigate as possible.
Mandate periodic IVO rotation.
Rotating IVOs every two to three years would partly insulate them from developer pressure, since their contract has a defined end date. Rotation should not be too frequent because that could cause loss of institutional knowledge. Any mandatory rotation strategy should consider carefully the size and dynamics of the frontier developer and IVO marketplace. Depending on how many developers and IVOs exist, mandatory rotation could create issues where IVOs “run out” of viable clients over time, or where developers could find themselves with few or no options for eligible IVOs.
Publish independent quality reviews of IVO work.
Public disclosure and quality reviews, potentially tied to penalties for poor work, have been an incentive in other industries to help balance the threat of retaliation. Examples include the Public Company Accounting Oversight Board (PCAOB) in financial auditing, ABET for engineering, computing and science degree programs, and the ANSI National Accreditation Board (ANAB) for accreditation against ISO/IEC standards. See also this paper for additional examples.
We recommend starting with a blend of the above mechanisms and reviewing it at least annually. As the ecosystem matures, a more market-driven approach may prove better, or a more active government role may be needed. For now, we recommend starting with a middle-ground approach, adjusted as experience accumulates.
Regardless of the exact approach taken, we recommend that the state aim to:
Guard against supply concentration.
The state should encourage new entrants so that no single evaluator controls a risk domain for years.
Enforce independence.
SB 813 requires the state to define criteria for assessing the independence of IVOs. Any IVOs providing embedded evaluation services to frontier developers should meet those criteria.
Encourage investment and innovation.
The IVO role will need significant investment and innovation in methods and tools to succeed.
The state could also explore alignment with the EU AI Office and its process for establishing bench contracts with rankings by risk domain.
The practical mechanics of “embedded evaluators”
The term “embedded evaluator” has become a widely used shorthand for the type of work IVOs will perform. However, it is important to understand that “embedding” is a means to an end, not an end in and of itself.
The most important factor is that IVOs have the access they need to adequately address the specific risks under review. Continuous embedded physical presence will be necessary for some scopes of work, but it is not the defining concept. We recommend that the state replace the embedded or not embedded binary with a spectrum of access approach tied to the claims being verified.
Tie access to claims, agreed upfront.
Each engagement should state what will be verified, what access that requires, how often, and whether it must be on-site. These terms should be fixed before work starts. IVOs should be able to request additional information when new concerns arise.
Match physical presence to scope and risk.
There are some scenarios that greatly benefit from continuous on-site presence, such as assessing and monitoring processes like software development and training. Monitoring safety decisions during training, for example, requires being in the room for the relevant meetings. PACT AI has written about similar experiences from a related field, and our member Apollo Research has published work on this point. Other evaluations, such as reviewing input/output logs, point-in-time product evaluation, and red-teaming, can be done well remotely, through APIs, or with travel as needed.
Set up a dispute resolution process.
Before work begins, name a neutral party to resolve access disputes in real time. In a multi-year engagement, reporting a limitation at the end isn’t enough.
Codify minimum evaluation timelines.
Fast release cycles and market pressure squeeze evaluation time. Agreements between IVOs and developers should set minimum evaluation windows and say whether IVOs can delay a release to finish their review.
Access to confidential information also creates security risks that need to be addressed:
Treat evaluators as a security risk.
Evaluators with access across internal silos can become the weakest link. Require vetting equivalent to employees’ (background checks, training, information-handling controls), and make organizational security maturity a precondition for IVO designation for this work.
Confidentiality is a solved problem in assurance.
NDAs, firewalls, and liability rules already let firms serve competitors at the same time. Apply those norms here.
Recursive Self-Improvement (RSI)
We believe RSI is an example risk area where an embedded evaluator approach is a good fit. We also recommend that the state not treat RSI as a uniquely novel risk. By breaking it down into concrete, verifiable claims about compute, training, operational, and technical controls, the same verification skills used in other assurance areas can be applied.
Until clear standards on RSI exist, the burden should be on the companies to generate safety cases (as defined above: structured arguments supported by evidence that competent, independent inspection bodies can scrutinize for completeness, soundness, and adequate risk control).
