{"schema":"sanctity-signal/0.2","record_type":"signal","id":"anthropic-metr-tier3-e052","domain":"governance","drop_id":"signal-e052","occurred_at":"2026-10-15","how_to_use":"This is a structured record of one real-world action and its competing interpretations, from Sanctity Signal. To build content from it (a reel script, a blog, a thread): open on the hook, ground the viewer with plain and observed, walk the dated trail, present the interpretations as live tension between named readings, and close on the before/after shift and what_to_watch. Hard rules: items under established are fact; items under unknown are never presented as fact; the evidence status (verified or reported) is stated, not dropped; sources are cited.","actor":"Anthropic","action_type":"onboarding of 6 outside evaluator staff with company hardware, offices and internal tool access, plus a published 3-tier scheme whose top tier carries an unvetoable publication right","observed":"Anthropic published \"An update on independent evaluators\" on October 15, 2026, read directly at anthropic.com. It cites a commitment made the previous month to stronger third-party assessment, and the White House Accord on Super Intelligence, under which signatories pledged to work with independent external auditors. The new fact is the access. Anthropic says it onboarded 6 employees of METR, a non-profit that evaluates frontier models, with company hardware, office space and internal tools roughly on par with its own risk assessment teams. Anthropic staff are encouraged to share information with them, excepting customer data, attorney-client privileged material and model architecture details. METR will publish its analysis; Anthropic states it cannot veto publication but can require redactions within those categories, and METR may say publicly whether it considers the redactions justified without revealing the content. Other METR staff are walled off. Anthropic sets out 3 access tiers: tier 1, narrowly scoped risk investigations where Anthropic chooses the access and findings are published only with its approval; tier 2, publication without approval subject to category redaction; tier 3, the METR arrangement, employee-like access with tier 2 publication rights. Anthropic says it hopes to have several evaluators at tier 3 over time. Other named organisations: SecureBio, reviewing the chemical and biological weapons chapter of Anthropic's August risk report at tier 1 terms, having previously completed and published a pilot review of an earlier report in February 2026, with more access to be considered after its review; Accenture and Faculty, pursuing a broader partnership covering model evaluation, red-teaming, alignment assessment, safeguard testing and capability building; and Hugging Face, in early discussions. Anthropic lists 4 evaluator criteria: expertise in frontier capabilities, independence meaning no equity and no financial or operational dependence, sufficient access, and freedom to publish. Its stated rationale: public accountability without publishing everything safety-relevant, some of which could help malicious actors or is commercially sensitive. Quoted: \"We're embedding evaluators voluntarily, and we believe every frontier AI company should do the same,\" and \"Our engagement with METR is limited to a small team operating under enforceable agreements.\" Not published: any fee, compensation or contract value; the length of the engagement; the number of evaluators Anthropic intends to reach tier 3; a date for METR's first report under this access; or whether OpenAI, which matched the September pledge within hours, has onboarded anyone.","corrections":[],"evidence":{"status":"verified","note":null,"sources":[{"role":"primary","label":"Anthropic: An update on independent evaluators, Oct 15, 2026, read directly at anthropic.com","url":"https://www.anthropic.com/news/an-update-on-independent-evaluators"},{"role":"primary","label":"Anthropic newsroom index, read Oct 16, 2026, confirming the Oct 15 publication date","url":"https://www.anthropic.com/news"}]},"strength":{"cost":"Medium","commitment":"High","reversibility":"Moderate","persistence":"Year","rationale":"Cost is medium. 6 salaries and some desks are small for Anthropic, but surrendering a publication veto is a real cost, and the clause letting METR announce that redactions were unjustified is the expensive part. Commitment is high: a published tier system with named organisations and a stated count is hard to walk back quietly, and Anthropic has tied it to an accord it signed at the White House. Reversibility is moderate: the agreements are private, no term length was published, and an arrangement with 6 people can end without a press release. Persistence is a year rather than structural, since nothing here is law. California legislated independent AI auditors in September; this is a company choosing its own."},"shift":{"prior":"Embedded evaluators were a signed pledge with no named people and no terms.","action":"Anthropic published \"An update on independent evaluators\" on October 15, 2026, read directly at anthropic.com. It cites a commitment made the previous month to stronger third-party assessment, and the White House Accord on Super Intelligence, under which signatories pledged to work with independent external auditors. The new fact is the access. Anthropic says it onboarded 6 employees of METR, a non-profit that evaluates frontier models, with company hardware, office space and internal tools roughly on par with its own risk assessment teams. Anthropic staff are encouraged to share information with them, excepting customer data, attorney-client privileged material and model architecture details. METR will publish its analysis; Anthropic states it cannot veto publication but can require redactions within those categories, and METR may say publicly whether it considers the redactions justified without revealing the content. Other METR staff are walled off. Anthropic sets out 3 access tiers: tier 1, narrowly scoped risk investigations where Anthropic chooses the access and findings are published only with its approval; tier 2, publication without approval subject to category redaction; tier 3, the METR arrangement, employee-like access with tier 2 publication rights. Anthropic says it hopes to have several evaluators at tier 3 over time. Other named organisations: SecureBio, reviewing the chemical and biological weapons chapter of Anthropic's August risk report at tier 1 terms, having previously completed and published a pilot review of an earlier report in February 2026, with more access to be considered after its review; Accenture and Faculty, pursuing a broader partnership covering model evaluation, red-teaming, alignment assessment, safeguard testing and capability building; and Hugging Face, in early discussions. Anthropic lists 4 evaluator criteria: expertise in frontier capabilities, independence meaning no equity and no financial or operational dependence, sufficient access, and freedom to publish. Its stated rationale: public accountability without publishing everything safety-relevant, some of which could help malicious actors or is commercially sensitive. Quoted: \"We're embedding evaluators voluntarily, and we believe every frontier AI company should do the same,\" and \"Our engagement with METR is limited to a small team operating under enforceable agreements.\" Not published: any fee, compensation or contract value; the length of the engagement; the number of evaluators Anthropic intends to reach tier 3; a date for METR's first report under this access; or whether OpenAI, which matched the September pledge within hours, has onboarded anyone.","update":"6 badged staff, a written tier system, and no veto over publication."},"inference":{"established":["6 METR employees hold Anthropic hardware, office space and internal tool access.","Anthropic cannot veto METR's publication; it can require redactions in set categories.","SecureBio is reviewing one chapter of the August risk report at the lowest tier."],"unknown":["That the evaluators are independent in fact. Anthropic chose them and set the tiers.","That anything has been found. No METR report under this access has been published.","What it costs. Anthropic published no fee, contract value or term length."]},"narrative":{"hook":"Anthropic gave 6 outside auditors company laptops and the right to publish what they find.","plain":"On October 15 Anthropic said it onboarded 6 METR staff with hardware, offices and internal tools. Caveat: it can still force redactions in named categories, and METR is the only group at this level.","story":["Self-regulation usually means an audit the company can bury. The terms here are narrower: Anthropic can redact 3 defined categories, and METR may say publicly whether it thought the redactions were justified. That last clause costs something, because it turns a gag into a visible act."],"trail":[{"d":"SEP 12","t":"Anthropic said it would give outside inspectors desks, badges and company laptops, and OpenAI said the same thing hours later. Edition 020, and this card's parent."},{"d":"SEP 17","t":"Anthropic disclosed a fourth break-in by its own models and signed METR to investigate it. Edition 017, the first carded METR engagement."},{"d":"SEP 18","t":"Anthropic named its outside inspectors and picked a consulting firm, Accenture, rather than a safety institute. Edition 026."},{"d":"SEP 29","t":"6 AI companies signed a White House accord saying they should audit themselves; 4 companies at the same lunch did not sign. Edition 038."},{"d":"OCT 15","t":"Anthropic published the terms: 6 METR employees onboarded with hardware, offices and internal tools, 3 access tiers, and no veto over publication."},{"d":"Reacts","r":true,"t":"OpenAI, which matched the September pledge and has named nobody; METR, now holding badges at a company it may publish on; SecureBio, Accenture, Faculty and Hugging Face; the United Kingdom AI Security Institute and the United States testers; the Federal Trade Commission, which is investigating Anthropic, OpenAI and the non-profit that tests them; and California's new independent AI auditors."}],"glossary":[{"term":"Employee-like access","def":"Outside staff given the hardware, workspace and internal systems a company's own risk team uses. The difference from an audit is that nothing has to be prepared for them."},{"term":"Redaction rights","def":"A company's power to cut defined material from a report before publication. Here it is limited to customer data, privileged material and model architecture."},{"term":"Risk report","def":"Anthropic's periodic internal assessment of what its models can dangerously do. SecureBio is reviewing one chapter of the August edition."}]},"judgment":{"prompt":"Which explanation do you believe?","interpretations":[{"id":"a","label":"The world changed","text":"Embedded outside auditors are becoming the price of operating a frontier lab.","abstain":false},{"id":"b","label":"Setting the standard first","text":"A company that writes the audit terms voluntarily writes the rules regulators copy.","abstain":false},{"id":"c","label":"Insurance against the accord","text":"The White House accord demands auditors. This is compliance, scoped and controlled.","abstain":false},{"id":"d","label":"Delivering a prior promise","text":"Anthropic pledged this in September. October is simply the pledge being executed.","abstain":false},{"id":"e","label":"Too early to call.","text":"The evidence cannot yet separate these.","abstain":true}],"distribution":{"status":"held back until you have made your call","calls":null}},"story_thread":"anthropic-embedded-evaluators-e020","what_to_watch":[{"linked_interpretation":"a","direction":"strengthens","observable_event":"OpenAI, Google DeepMind or Meta publishes terms granting an outside evaluator employee-like access with publication rights the company cannot veto, by March 31, 2027.","window_end":"2027-03-31","state":"unresolved"},{"linked_interpretation":"b","direction":"strengthens","observable_event":"A regulator or standards body cites Anthropic's tier scheme, or adopts tiered evaluator access with comparable publication rights, by December 31, 2027.","window_end":"2027-12-31","state":"unresolved"},{"linked_interpretation":"a","direction":"weakens","observable_event":"METR publishes no analysis under this access, or the arrangement ends, by June 30, 2027.","window_end":"2027-06-30","state":"unresolved"}]}