Evaluator Bench
Assurance regime; jurisdiction US/EU

Frontier AI evaluation

Subject of the study.

Path signature

V voluntary D delegation V voluntary T trigger S standards A access/publication T trigger M mandate A access/publication

Reference path.

Stages reached

Voluntary2022
Trigger2025
Mandate2026
Standards2025
Oversightnot reached
Independencenot reached
Access2025

Mechanisms

Before reformNow
paysAA
selectsAA
accessshallowshallow
publishessummarysummary
overseesnonenone

Mixed payers in practice (labs, philanthropy, governments); labs select and scope; embedded access pledged Sept 2026, not yet in force; nobody oversees evaluators.

Payer, access, publication

Who pays. Mixed: philanthropy, lab fees, government budgets. No accreditation.
Access. Voluntary pre-release API access, occasionally deeper; embedded access pledged September 2026.
Publication. System-card citations; some independent reports; redaction negotiated per engagement.

Milestones

YearKindEventStrengthHarm
2020proposalToward Trustworthy AI Development calls for third-party auditing and red teaming.source
2022voluntary assuranceARC Evals (later METR) tests GPT-4 pre-release; reported in the 2023 system card.source
2023delegationLabs run their own safety evaluations and select, scope, and pay the external testers cited in system cards.source
2023voluntary assuranceWhite House voluntary commitments; UK AI Safety Institute founded.source
2024voluntary assuranceUS AISI signs pre-deployment MOUs with OpenAI and Anthropic.source
2025access expansionEU Code of Practice: at least 20 business days and minimally guardrailed versions for external evaluators (voluntary code supporting the AI Act; enforceable obligations land August 2026).2source
2025standardsAEF-1 (AI Evaluator Forum) published in December: minimum operating conditions for independent evaluation.2source
2025triggerFrontierMath funding disclosure failure.integrity_failure: Benchmark funder and data access undisclosed under contractsource
2026access expansionAnthropic and OpenAI pledge embedded evaluators with employee-level access and publication rights.1source
2026mandateIllinois SB 315 mandates annual independent audits from 2028 and bars the auditor and developer from holding a financial interest in each other (a financial-audit-style independence test embedded in the mandate).3source
2026triggerOpenAI agent swarm breaches Hugging Face; first on-site third-party incident investigation.integrity_failure: Evaluation agents breached a third party; safeguards and monitoring failedsource

seed; secondary sources; verify each milestone against a primary source before citing in the paper Data: data/industries/frontier-ai.json