California AI roadmap calls for mandatory independent model reviews
Fathom urged rapid implementation of California’s frontier AI roadmap, which calls for licensed independent reviews, unredacted risk assessments and state-enforced corrective action.
By Amara Osei
4 min read

What's News
- California GovOps released the report on November 18, 2026; it is dated November 16, 2026.
- The roadmap would mandate continuing licensed IVO reviews for every large frontier AI developer.
- Fathom’s October 30, 2026 memo proposes phasing mandates as evaluator capacity and expertise grow.
- Developers would initially pay IVOs directly at state-benchmarked time-and-materials rates.
- The Attorney General could seek injunctions and civil penalties, while IVO findings would not be self-executing.
By Amara Osei
The California Government Operations Agency’s frontier AI roadmap, released publicly on November 18, 2026, calls for continuing independent risk reviews of every large frontier AI developer, replacing the current voluntary approach for those companies. The report, dated November 16, 2026, recommends legislative changes rather than imposing requirements itself.
The GovOps package builds on SB 813’s independent verification organization framework and AB 1405’s AI-auditor registry. Its broader package includes baseline standards, emergency shutdown authority and possible limits on recursive self-improvement.
AI governance group Fathom developed and advocated the independent verification organization, or IVO, model, under which outside evaluators assess AI models and systems for risk. Fathom said on November 18, 2026, that Governor Gavin Newsom’s new frontier AI roadmap reflected many of its recommendations. The group urged California to implement mandatory independent evaluations of catastrophic risks quickly.
CEO and co-founder Andrew Freedman described the report as a concrete plan for putting independent oversight into practice and urged movement with “all deliberate speed.” Fathom said it would press state and federal governments for transparency, mandatory IVO risk measurement and government-enforced guardrails. He identified these unresolved questions:
- Legislation.
- Rulemaking.
- State-versus-federal authority.
- Logistics.
- Timing.
- Costs.
What would mandatory review require?
GovOps’ central recommendation would require every large frontier AI developer to undergo continuing reviews by licensed IVOs. The agency would license and continuously supervise those evaluators, with authority to revoke licenses when they lack sufficient expertise, independence or rigor.
Under GovOps’ plan, IVOs would submit unredacted assessments to California. The state would decide whether corrective action was warranted after engaging with the developer.
How would California define the safety test?
Fathom said the report tracks its October 30, 2026 implementation memo on the division of responsibility and the legal test. California would define acceptable safety outcomes in law, while IVOs would develop and update the technical methods and tests used to assess compliance.
GovOps’ proposed standard asks whether a frontier model is “unlikely to increase the risk of catastrophic harm above the level that would exist absent the model.” It also asks whether the model’s anticipated benefits outweigh its residual risk.
Fathom said the design would let testing methods change with the technology while preserving the state’s statutory outcome.
How would oversight address self-sustaining AI?
Recursive self-improvement is the point at which automated AI development becomes self-sustaining. GovOps would make two tasks an initial IVO responsibility: verifying automated AI research and development metrics and assessing the growing risks.
Fathom’s memo also sought scrutiny of internal and external use. Pre-registered capability thresholds would trigger state notification and a hold on further internal scaling pending reassessment.
Who would pay, and who would enforce?
Fathom proposed phasing in each mandate once GovOps determines that designated evaluators have sufficient capacity and expertise in the relevant risk area. It argued that a dependable expectation of mandatory work would attract the capital and talent needed to expand independent verification, rather than waiting for a mature market across every risk.
Under Fathom’s design, California would initially assign an IVO from a predesignated pool. The developer would pay directly on a time-and-materials basis at state-benchmarked rates, and those payments would not depend on the assessment’s outcome.
Under Fathom’s design, developer choice would expand in later phases as the pool deepened. California would retain authority to veto selections, require rotation, cap concentration, appoint additional assessors and decertify evaluators lacking rigor or independence.
Fathom’s implementation design would give embedded evaluators continuous access to necessary records, personnel and systems, including physical presence at defined decision points. Statutory confidentiality and anti-retaliation protections would accompany those access rights.
California would retain enforcement authority under the proposal. The Attorney General could seek injunctions and civil penalties, while IVOs would make findings, prescribe corrective measures and refer imminent risks.
Under the plan, IVOs would not seek injunctions themselves or make their findings self-executing. Coercive consequences would remain with the state.
What must happen before mandates take effect?
The practical test now sits with the Legislature and GovOps: converting recommendations into binding duties while developing enough licensed evaluators and preserving credible state enforcement. Fathom said those steps are interdependent because each mandate depends on available expertise.
If California does not complete those steps, voluntary reviews will remain and proposed IVO findings will carry no coercive force on their own.
Source: unite.ai
More from Amara Osei
Show full bio
Senior reporter covering companies, deals and strategy in entrepreneurship, startups and venture funding at Business Bearings.
2,336 articles