The weirdest job in AI may soon be reading someone else's model safety homework and saying, politely, prove it. California's SB 813 is not about another leaderboard where Model A beats Model B by a suspicious decimal point. It is about who gets to check frontier AI safety claims, under what authority, and with what incentives. That last bit matters, because an audit funded like a party favor is not quite the same thing as an audit funded like public infrastructure. Fathom, the nonprofit that sponsored SB 813, says the California Legislature overwhelmingly passed the bill authored by Senator Jerry McNerney. Fathom describes the measure as a step toward the country's first state framework for Independent Verification Organizations, or IVOs, to test AI safety claims, measure risk, and build public trust. Translation from policy dialect: the state is trying to make model evaluation less like a self graded take home exam and more like an inspection regime with adult supervision. AI labs may hate the paperwork, but paperwork is how societies tell very large machines to stop vibes based deployment. ## SB 813 moves AI auditing from vibes to institutions According to Fathom, the core mechanism in SB 813 is the IVO model, outside experts that can evaluate AI safety claims and risk. ICT Solutions and Education reports that California lawmakers passed SB 813 to create a framework for independent organizations to assess AI safety, risk, and safeguards. That wording is important because it points beyond benchmark theater. A credible verifier is not just asking whether a model can ace a test set, it is asking whether the developer's safety story survives contact with independent scrutiny. CalMatters Digital Democracy notes that existing law, the Transparency in Frontier Artificial Intelligence Act, already requires a large frontier developer to write, implement, and clearly publish a frontier AI framework for its frontier models. SB 813 fits naturally after that: if developers must publish frameworks, someone outside the developer needs a way to inspect whether those frameworks mean anything. Otherwise, a safety framework can become the AI equivalent of a gym membership bought in January and ignored by February. ## GovBuddy shows the policy plumbing is now moving GovBuddy lists SB 813 in the California 2025 to 2026 regular session and identifies it as a Senate bill with the status Passed both house. That does not make the technical questions disappear. It makes them operational. Once verification becomes a state recognized function, labs, evaluators, and public agencies need boring but essential answers about scope, access, documentation, and conflict management. This is where the engineering rubber meets the regulatory road, then complains that the road was not in the requirements doc. A model evaluator may need API access, system cards, policy documentation, threat models, evaluation harnesses, and enough compute budget to run meaningful tests. If the evaluator cannot afford the tests, independence becomes decorative. If the developer controls the access too tightly, independence becomes dinner theater with JSON outputs. ## The money problem is not a footnote ICT Solutions and Education frames IVOs as qualified third party assessments of AI models and systems, including safety, risk, and safeguards. That sounds clean, but anyone who has run serious model evaluations knows the meter starts spinning quickly. Long context probes, agentic task suites, adversarial testing, repeated sampling, tool use evaluation, and regression checks are not free. The invoice is not the whole story, but pretending it is irrelevant is how we get audits that are basically PDFs wearing lab coats. The policy lesson is simple: verifier independence needs a funding model, not just a mission statement. If labs pay directly, readers should ask how conflicts are managed. If the state pays or certifies a market of evaluators, readers should ask how quality and access are maintained. If nobody pays enough, the ecosystem gets symbolic testing, which is like checking a parachute by complimenting the stitching. ## What builders should prepare for next For AI builders, the practical takeaway is not panic, it is hygiene. Fathom says SB 813 is designed around testing safety claims and measuring risk, so teams should make those claims testable before someone else asks for evidence. That means versioned model documentation, evaluation logs, incident response procedures, clear release criteria, and a record of what changed between model versions. Yes, this is less glamorous than a benchmark victory lap, but so is flossing, and dentists remain undefeated. For product and safety teams, the emerging market for independent verification means release readiness may start to look more like compliance plus engineering review. The best teams will treat this as feedback infrastructure, not merely a permission slip. Watch how California defines qualified evaluators, how access terms are structured, and whether funding models support real independence. The audit economy is arriving, and it is bringing receipts. ## Sources - California Legislature Overwhelmingly Passes Fathom-Sponsored Bill to Spur Independent Verification of AI Safety - Fathom

Sources