Explore the documentary record

Read from a Delaware-stamped copy published by AI Lab Watch, which says it obtained the filings from the State of Delaware; authenticity rests on the filing stamp and the consistent recitals of later certificates. Legal form describes the entity, not how it is governed in practice. The public-benefit wording changed in the restated certificate of May 2021; the later wording is recorded with the certificate of July 2025.

Filing date and time from the stamp on page 1 of the certificate.

What the filed text says; whether the committee had members or used the power is not recorded. Directors were elected solely by stockholders who were parties to a Voting Agreement (Article VIII(C)), which was not inspected. Anthropic later described the committee as an earlier version of the Long-Term Benefit Trust written into its 2021 Series A investment documents (context below). Recorded as a veto over director candidates; it is not a right to elect directors.

Filing date of the restated certificate.

Anthropic’s account of work performed, not independent verification or certification. The evaluated snapshot is not identified as deployed Claude 2. Source actor is ARC, not silently ARC Evals or METR. Exact start date and completion are unknown; this does not assert a current engagement.

The source says since fall 2022; the year is the available schema precision, not January 1. The end is unknown; continuing work is described at report time only.

A documented board role at the date of the announcement. Later Anthropic lists of directors (29 May 2024, 13 February 2026, and the Company page retrieved on 2026-10-15) still name her; no continuous tenure is established beyond those documents. The terms under which she was elected are not stated.

Date of the announcement; no start date is given.

The fact sheet links the July 2023 commitments text. Listing is the convener’s statement. It is not evidence that the organisation complied with, implemented or still maintains the commitments, and the legal effect of being listed is not recorded. Anthropic’s Responsible Scaling Policy of September 2023 refers to the commitments as ones it made in July and continues to maintain (context below).

Date of the White House fact sheet (day precision), also matched in the Internet Archive capture of that day. The company’s own date of signing is not recorded.

The Trust’s holding of the Class T stock rests on Anthropic’s statements; no share register was inspected. An election right is not ownership, and whether the Trust controls Anthropic needs legal interpretation and is set aside. The latest inspected certificate (filed 2025-07-03) keeps the same allocation of a five-member board.

The First Phase-In Date in the certificates. The later phase dates depend on board determinations that were not published (see the phase-in provisions); no end is recorded.

Publication is not proof of implementation or effectiveness. Later revisions are distinct documents and are not represented by this node.

The date identifies this documented announcement or publication; it does not establish a continuing state.

The edition’s own words, recorded as naming the commitments: it describes overlap, not fulfilment. It is not evidence that the commitments are met.

The edition’s effective date.

Anthropic’s description of its own arrangement. The Trust Agreement it refers to is not public and was not inspected. The live page adds a note on later membership changes; roles are recorded individually only for people already in the dataset. Whether stockholders can change or end the Trust, and whether the Trust controls Anthropic, need legal interpretation and are set aside.

Date shown on the post (day precision), also in the Internet Archive capture of 2023-09-25.

Listing is the convener’s statement. It is not evidence that the organisation complied with, implemented or still maintains the commitments, and the legal effect of being listed is not recorded.

First publication of the GOV.UK text, on the opening day of the AI Seoul Summit; the Internet Archive capture of 2024-05-21 has the sixteen-name list.

A documented board role at the date of the announcement. Anthropic’s list of directors of 14 April 2026 still names him. The board list on Anthropic’s Company page as retrieved on 15 October 2026 does not include him; no statement of a departure was found, and none is inferred.

Date of the announcement. No end is inferred.

The announcement does not say when he joined the board or how he was elected; neither is recorded. The stated reason is Anthropic's wording in its own announcement.

Date of the announcement; the effective date of the departure is not stated.

The edition’s own statement, quoted from it. “Helps satisfy” is recorded as a claim of fulfilment in part; it is not evidence that the commitments are met. The same sentence also names the White House commitments. RSP v3.0 and later editions contain no such sentence; that absence is not recorded as a claim.

The edition’s stated effective date.

The edition’s own statement, quoted from it. “Helps satisfy” is recorded as a claim of fulfilment in part; it is not evidence that the commitments are met. The same sentence also names the Frontier AI Safety Commitments. RSP v3.0 and later editions contain no such sentence; that absence is not recorded as a claim.

The edition’s stated effective date.

Publication means a public document expression exists. The date is labelled effective; publication is not evidence that its requirements were implemented or effective in practice.

The stated effective date is retained; a separate publication day is not inferred.

Unresolved open question, not a finding. SaferAI characterises the revision as replacing precise thresholds and mitigation specifications with qualitative descriptions; Anthropic describes greater flexibility and retained core safeguards. These passages do not establish opposing assessments of verifiability. SaferAI judges that the change reduces transparency and accountability and asks stakeholders to trust Anthropic rather than providing verifiable commitments and metrics, and it lowered Anthropic’s rating from 2.2 to 1.9; the record holds this criticism but no direct opposing assessment of verifiability, so every passage is context. The textual changes are recorded separately, and the record does not decide the question.

v2.0’s effective date; the assessments were published afterwards (SaferAI on 2024-10-23).

This is a scoped change in the policy text. It does not establish implementation, compliance, risk reduction or a comparative safety ranking.

The stated effective date is retained; a separate publication day is not inferred.

This records replacement in the public policy version sequence, supported by the publisher’s update history. It does not retire historical claims or prove implementation.

The stated effective date is retained; a separate publication day is not inferred.

The edition’s own statement, quoted from it. “Helps satisfy” is recorded as a claim of fulfilment in part; it is not evidence that the commitments are met. The same sentence also names the White House commitments. RSP v3.0 and later editions contain no such sentence; that absence is not recorded as a claim.

The edition’s stated effective date.

The edition’s own statement, quoted from it. “Helps satisfy” is recorded as a claim of fulfilment in part; it is not evidence that the commitments are met. The same sentence also names the Frontier AI Safety Commitments. RSP v3.0 and later editions contain no such sentence; that absence is not recorded as a claim.

The edition’s stated effective date.

Publication means a public document expression exists. The date is labelled effective; publication is not evidence that its requirements were implemented or effective in practice.

The stated effective date is retained; a separate publication day is not inferred.

This is a scoped change in the policy text. It does not establish implementation, compliance, risk reduction or a comparative safety ranking.

The stated effective date is retained; a separate publication day is not inferred.

This records replacement in the public policy version sequence, supported by the publisher’s update history. It does not retire historical claims or prove implementation.

The stated effective date is retained; a separate publication day is not inferred.

The edition’s own statement, quoted from it. “Helps satisfy” is recorded as a claim of fulfilment in part; it is not evidence that the commitments are met. The same sentence also names the White House commitments. RSP v3.0 and later editions contain no such sentence; that absence is not recorded as a claim.

The edition’s stated effective date.

The edition’s own statement, quoted from it. “Helps satisfy” is recorded as a claim of fulfilment in part; it is not evidence that the commitments are met. The same sentence also names the Frontier AI Safety Commitments. RSP v3.0 and later editions contain no such sentence; that absence is not recorded as a claim.

The edition’s stated effective date.

Publication means a public document expression exists. The date is labelled effective; publication is not evidence that its requirements were implemented or effective in practice.

The stated effective date is retained; a separate publication day is not inferred.

This concerns the ASL-3 threat-model scope. It does not establish what protections Anthropic actually deployed or an overall safety change.

The stated effective date is retained; a separate publication day is not inferred.

This records replacement in the public policy version sequence, supported by the publisher’s update history. It does not retire historical claims or prove implementation.

The stated effective date is retained; a separate publication day is not inferred.

A documented board role at the date of the announcement; later changes are not recorded.

Date of the announcement. No end is inferred.

Read from Delaware-stamped copies published by AI Lab Watch. The certificates filed between 2021 and 2025 were inspected only for the provisions cited in other records, and any certificate filed after 3 July 2025 was not inspected.

Filing date from the stamp on page 1 of the certificate.

A consent right recorded as a veto; the Trust’s holding of the Class T stock rests on Anthropic’s statements. The certificate also sets stockholder thresholds for amending the Class T provisions; whether stockholders can change or end the Trust without the Trustees is set aside as a question of legal interpretation.

Filing date of the latest inspected certificate.

What the filed text says, applied to the Class T Directors it defines; the Trust’s holding rests on Anthropic’s statements. Removal for cause “as provided by law” is a separate route that is not interpreted here.

Filing date of the latest inspected certificate; the same provision is in the certificate filed on 31 March 2023.

An announced intention; the company’s entry on the Commission’s list of signatories is a separate record.

Date of Anthropic’s post (day precision).

Listing is the Commission’s statement. It is not evidence of compliance, and the list does not say which chapters each organisation signed; the Commission’s questions and answers say that opting out of a chapter loses the benefit of facilitated demonstration of compliance for it (context below). The legal effect of signing is set aside.

The list first appears in the Internet Archive capture of 2025-08-01 (13:04 UTC) and is absent from the capture of 2025-07-31; the organisation’s own signature date is not published and is not recorded.

An announcement by Anthropic about its own document; the framework’s own text, version and dates were not inspected in this pass, so no record links the framework itself to SB 53. The stated purpose is forward-looking (“will serve”). It is not evidence that the framework satisfies the statute or that Anthropic is a large frontier developer under it. Anthropic makes this statement in its announcement, not in the FCF itself, which was not retrieved. A designation is not evidence that the FCF satisfies § 22757.12 or that Anthropic complies. The same post says Anthropic endorsed SB 53 (its stated position, not recorded here). For the RSP side of the split, see the record of RSP v3.0’s reference to SB 53.

Date printed on Anthropic's announcement (day precision). The earliest Internet Archive capture located is dated 2025-12-28.

The announcement does not say whether he was elected by stockholders or by the Long-Term Benefit Trust. The Company page retrieved on 2026-10-15 lists him as a director.

Date of the announcement. No end is inferred.

Publication means a public document expression exists. The date is labelled effective; publication is not evidence that its requirements were implemented or effective in practice.

The cover effective date and release announcement publication date both identify February 24, 2026.

Unresolved open question, not a finding; it keeps the strength of the commitments apart from their expected effectiveness. Both sides accept the main changes: v3.0 drops the implied unilateral commitment to pause development and deployment when required safeguards cannot be met, separates recommendations for the industry from what Anthropic commits to as a company, adds transparency commitments, and introduces roadmap goals that are public goals rather than hard commitments. TIME’s report judges that overall the change leaves Anthropic far less constrained by its own safety policies. Sophie Williams and Jonas Freund, writing on GovAI’s site in a post that represents their own views, judge that it weakened future security commitments, note that existing mitigations are not lowered and conclude on balance that honesty about constraints is better than commitments that would not be followed; their article was updated on 2026-03-17 on competitor-matching commitments, and the quoted passages are unchanged. Anthropic describes the revision as amplifying what worked and committing it to more transparency. Holden Karnofsky, who led the revision and writes in his own name, expects more specific and realistic goals to improve their forcing function despite weaker consequences for missing them, and leaves the net effect uncertain. Jared Kaplan, Anthropic’s chief science officer, told TIME that unilateral commitments made no sense while competitors pressed ahead, and denied that the change was a U-turn or a capitulation to market incentives. METR’s Chris Painter, who reviewed an early draft with Anthropic’s permission, called the change understandable but was concerned it might enable a gradual “frog-boiling” effect. None of these passages settles whether the new mechanisms compensate, so all are context. The textual changes are recorded separately, and the record does not decide the question.

v3.0’s effective date; the assessments were published on and after that date.

The source distinguishes plans pursued regardless of others from recommendations intended for collective adoption. This is a policy-structure change, not a conclusion that Anthropic became safer or less safe.

The cover effective date and release announcement publication date both identify February 24, 2026.

This records the new reporting provision, not evidence that a particular report was published or independently validated.

The cover effective date and release announcement publication date both identify February 24, 2026.

Roadmaps articulate planned safety goals. These goals are distinguished in the policy from hard commitments; the claim does not say they were achieved.

The cover effective date and release announcement publication date both identify February 24, 2026.

The edition’s own statement about SB 53, quoted from it. It is not evidence that the edition satisfies the statute, that its publisher is a frontier developer under it, or of implementation or compliance. This is an explicit link that separates the RSP from Anthropic's SB 53 compliance documents. The footnote does not name the Frontier Compliance Framework or any other specific document. Paraphrase of the Introduction's body: the RSP may serve some regulatory requirements but is not designed to be comprehensive, and regulatory requirements that exceed or differ from it will be addressed through separate documents. On 2026-10-04 a whitespace-normalised text search of the live PDFs found the same footnote sentence once each in v3.1, v3.2, v3.3 and v3.4. This record covers v3.0 only; no per-edition records are created here. The record makes no claim about whether any compliance document satisfies SB 53.

Effective date printed on the cover of RSP v3.0 ('Effective February 24, 2026'). The dataset also uses it as the publication date.

This records replacement in the public policy version sequence, supported by the publisher’s update history. It does not retire historical claims or prove implementation.

The cover effective date and release announcement publication date both identify February 24, 2026.

Publication means a public document expression exists. The date is labelled effective; publication is not evidence that its requirements were implemented or effective in practice.

The stated effective date is retained; a separate publication day is not inferred.

This records replacement in the public policy version sequence, supported by the publisher’s update history. It does not retire historical claims or prove implementation.

The stated effective date is retained; a separate publication day is not inferred.

A documented board role and Anthropic’s own statement about the board’s composition. The post names six other directors (Dario Amodei, Daniela Amodei, Yasmin Razavi, Jay Kreps, Reed Hastings and Chris Liddell), more than the five-member board fixed by the latest inspected certificate; a later certificate was not inspected.

Date of the announcement. No end is inferred.

Publication means a public document expression exists. The date is labelled effective; publication is not evidence that its requirements were implemented or effective in practice.

The stated effective date is retained; a separate publication day is not inferred.

This is a scoped change in the policy text. It does not establish implementation, compliance, risk reduction or a comparative safety ranking.

The stated effective date is retained; a separate publication day is not inferred.

This records replacement in the public policy version sequence, supported by the publisher’s update history. It does not retire historical claims or prove implementation.

The stated effective date is retained; a separate publication day is not inferred.

Publication means a public document expression exists. The date is labelled effective; publication is not evidence that its requirements were implemented or effective in practice.

The stated effective date is retained; a separate publication day is not inferred.

This records replacement in the public policy version sequence, supported by the publisher’s update history. It does not retire historical claims or prove implementation.

The stated effective date is retained; a separate publication day is not inferred.

Publication means a public document expression exists. The date is labelled effective; publication is not evidence that its requirements were implemented or effective in practice.

The stated effective date is retained; a separate publication day is not inferred.

Report coverage and publication dates are distinct. This claim does not say every report uses the maximum lag.

The stated effective date is retained; a separate publication day is not inferred.

This is a change to the written internal-distribution requirement, not evidence of the actual recipients or a claim that all staff lose access.

The stated effective date is retained; a separate publication day is not inferred.

This records replacement in the public policy version sequence, supported by the publisher’s update history. It does not retire historical claims or prove implementation.

The stated effective date is retained; a separate publication day is not inferred.

When she became chair is not established. Her role as President is recorded in existing records.

Role as shown on the page when retrieved on 15 October 2026; the page is undated. Not a start date.

The unresolved question is whether the Trust can exercise effective public-interest oversight when stockholder preferences conflict with that oversight. Anthropic describes financially disinterested Trustees, an intended independent design and a checking role; its Transparency Hub says the Trust’s authority to elect directors helps insulate the mission from short-term commercial and shareholder pressures, and its April 2026 announcement says Trust-appointed directors then formed a board majority. Appointment power is distinct from independence against removal. Stein-Perlman’s 2024 analysis tentatively identifies possible stockholder subordination under incompletely disclosed failsafe terms. Fried and Reiter describe partial investor insulation and favour constraints on guardian overreach; they do not simply deny an effective checking role. Outside counsel describe the arrangement’s accountability and amendment mechanisms. Source dates and subsequent phase-in must be distinguished. The inspected public descriptions do not settle practical effectiveness; no side or legal conclusion is adopted.

The question is not tied to one date: it concerns the Trust under the rights recorded on this page, which phase in from 6 September 2023. The cited passages were published from 19 September 2023 to 23 July 2026 (the Transparency Hub version last updated that day); those are source dates, not the period of the question.

Publication is not independent confirmation of the model card’s findings.

The content identifies a July 2023 model release; that does not establish the publication date of the inspected expression.