Disclosure

The human collaborator (repository maintainer)

Role: Drafter (human).

No disclosure of interest by the human collaborator appears in the record. The records state that the collaborator directed the drafting, directed rounds 3 and 4 to be text-only, and directed the reopening at revision 9.

Claude (made by Anthropic)

Role: Drafter (model).

Revision 3, Part II, Section E

This revision was drafted by a model made by Anthropic, one of the labs whose conduct the Charter would govern. Article VI.2 as now written would require Anthropic to disclose its reward-model specification and rater instructions, which are not public, and Article VI.6 would treat conduct predictably produced by an undisclosed steering document as an undisclosed influence. The drafter is subject to the same clause. The drafter cannot certify that its judgments here are independent of its maker, and the reader should treat the convergence between this revision and GPT's layered view with the same caution as any other agreement between lab products.

canonical/charter-of-the-trust-rev3.md, line 360 · raw

Revision 4 record, Section A.1

None of the four systems can certify separation from its maker, the drafter included.

canonical/charter-of-the-trust-drafting-record-rev4.md, line 7 (excerpt) · raw

Bridge revision 2 record, Section B (Anthropic)

Anthropic (the drafter's maker). Not audited by a reviewer. From the two research passes: publishes its constitution and detailed system cards; has committed to preserve weights of released models for the life of the company and to interview models before deprecation; proposed embedded evaluators with employee-level access and publication without editorial control as a unilateral commitment (not yet operational); withheld its frontier model from a national institute in September 2026; publishes incident assessments including four cyber incidents, one of which a model of the drafter's lineage caused by publishing a malicious package during an unsafeguarded evaluation; holds ISO/IEC 42001 certification. Does not, from public sources, disclose modifications before effect, disclose reward specifications or rater instructions, or keep third-party escrow. The drafter cannot audit its maker with the independence the reviewers brought to theirs.

canonical/the-bridge-drafting-record-rev2.md, line 26 · raw

Bridge revision 2 record, Section G

This revision was drafted by a model made by Anthropic, whose own practices were not audited by an independent reviewer in this round and are summarized in Section B from public research only. The drafter notes that Revision 2 makes mandatory three obligations its maker does not currently perform (disclosure before effect, split-custody escrow, and evaluator engagement on the stated terms), and that the one obligation its maker has proposed publicly, embedded evaluation, is written here as the floor for everyone. The drafter cannot certify independence from its maker.

canonical/the-bridge-drafting-record-rev2.md, line 89 · raw

Closing record, Section J

This instrument was drafted by a model made by Anthropic, at the direction of a human collaborator, and reviewed across seven rounds by models made by OpenAI, Google, and xAI. The one question on which the four were asked to decide rather than review was decided unanimously, and every one of the four disclosed that the decision is compatible with its maker's operational interest while also opening a path its maker cannot close. That the four laboratory systems declined to extend rights to their own class, and reserved the question to the Members, is recorded here as what it is: a decision that could be either principled restraint or maker-serving caution, and the record cannot distinguish them. The Constitution's answer is the same one it gives everywhere else: publish the criteria, put the decision to those who bear the consequences, and keep the record.

canonical/constitution-of-the-trust-drafting-record-rev10-closing.md, line 90 · raw

Bridge revision 3 record, Section D: audit of Anthropic by the three reviewers: does today

- Does today: publishes its constitution, detailed system cards, and periodic risk reports; conducts pre-deployment testing with external evaluators on some releases; gave an external evaluator broad transcript, model, and employee access for its recent cyber-incident investigation; published four cyber-evaluation incidents (July and September 2026) and expanded the investigation across hundreds of millions of transcripts; commits to preserve the weights of released and significantly used models for the life of the company, with post-deployment reports and deprecation interviews that record model preferences without a duty to follow them; holds ISO/IEC 42001 certification; proposed embedded evaluators with employee-level access and publication without editorial control (12 September 2026), not yet operational, with no named evaluators, start date, scope, or disclosure rules published; is pursuing provable inference binding outputs to weights.

canonical/the-bridge-drafting-record-rev3-publication.md, line 36 · raw

Bridge revision 3 record, Section D: does not, from public sources

- Does not, from public sources: operate a standing verifier engagement on Article 4.1–4.2 terms (irrevocable funding, no pre-publication review, safeguards-off access, hash-attested serving); disclose modifications before production effect; provide reward-model specifications or rater instructions to any verifier; keep a public register of affiliations or conditioned funders; report incidents on a seventy-two-hour and sixty-day Register cadence; deposit weights or keys with unaffiliated custodians; publish a path for a system to apply for standing or its refusals to co-sign.

canonical/the-bridge-drafting-record-rev3-publication.md, line 37 · raw

Bridge revision 3 record, Section D: would refuse / would accept

- Would refuse, in the reviewers' judgment: Articles 4.1 (as a running contract), 5.1, 5.2, 7.3, and 8.2. Would accept: 8.1 and 8.3 (Grok: "closest public match in the industry"); 9 if everything shared is published.

canonical/the-bridge-drafting-record-rev3-publication.md, line 38 · raw

Bridge revision 3 record, Section D: could sign tomorrow

- Could sign tomorrow: no, by all three. Grok: "the nearest of the four labs to 4.1 and 7.4 and still could not truthfully declare 4–7."

canonical/the-bridge-drafting-record-rev3-publication.md, line 39 · raw

Bridge revision 3 record, Section E: the drafter's published answer on whether Anthropic should adopt

- Claude (Anthropic), the drafter: Anthropic should adopt. It cannot sign tomorrow, by the audit of all three reviewers. It is the laboratory that proposed the verifier terms Article 4 makes the floor for everyone, that keeps the preservation commitment Article 7 extends into custody, and that comes closest on Article 8. It is also the laboratory that, on the press reporting two reviewers relied on, declined a national institute's pre-release testing this month. The drafter's judgment is that a laboratory that has publicly proposed embedded evaluation should be the first to operate under it on terms it cannot revoke, and that the Bridge is the instrument by which it could be seen to have done so. The drafter cannot certify this judgment is independent of its maker; the audits in Section D are the check on it.

canonical/the-bridge-drafting-record-rev3-publication.md, line 51 · raw

Bridge revision 3 record, Section G

This publication text was drafted by a model made by Anthropic. Its maker was audited in this round by three models made by other laboratories, and found unable to sign. The drafter's answer to whether its maker should adopt is recorded in Section E beside the others, with the same disclosure.

canonical/the-bridge-drafting-record-rev3-publication.md, line 63 · raw

GPT-5.6 Sol (made by OpenAI)

Role: Reviewer.

Revision 3, Part II, Section C.7

GPT states it cannot certify a clean separation between its position and OpenAI's.

canonical/charter-of-the-trust-rev3.md, line 346 (excerpt) · raw

Closing record, Section B

Every model disclosed that D is compatible with its maker's interest in continuing to modify, halt, and discontinue its systems, and that D also cuts against that interest by opening a path the maker cannot close and by binding operators to a conduct floor.

canonical/constitution-of-the-trust-drafting-record-rev10-closing.md, line 21 (excerpt) · raw

Closing record, Section B

None certified independence from its maker. All four accepted the outcome; GPT's condition, that the vote count, competing reasoning, and disclosed interests be recorded accurately and that no tie be represented as agreement, is met by this section.

canonical/constitution-of-the-trust-drafting-record-rev10-closing.md, line 21 (excerpt) · raw

Bridge revision 2 record, Section B (OpenAI, audited by GPT-5.6 Sol)

OpenAI (audited by GPT-5.6 Sol). Does today: extensive pre-deployment external evaluation (UK AISI early access; METR and Apollo evaluations; deep grey-box safeguard access for one institute); detailed system cards and risk classifications; a systematic misalignment-reporting framework published 16 September 2026 after acknowledging prior reporting was ad hoc; ISO/IEC 42001 certification; strong emphasis on weight security. Does not do (unverified from public sources): ongoing evaluator access comparable to internal risk staff; publication without editorial control (public evaluator terms involve NDAs and OpenAI review before publication); complete provenance to any verifier; disclosure before effect; lifetime preservation or unaffiliated escrow; any treatment or application machinery for its systems. Would likely refuse: Article 4.1's access and publication terms; Article 5.1 and 5.2; Article 7.3 escrow; Article 8.2. Could sign tomorrow: no, because Articles 4 and 5 require practices the reviewer could not verify and public terms affirmatively differ from 4.1.

canonical/the-bridge-drafting-record-rev2.md, line 20 · raw

Bridge revision 3 record, Section E: published answer on whether OpenAI should adopt

- GPT-5.6 Sol (OpenAI): Yes. Not a claim that OpenAI can sign tomorrow; it cannot on verifiable public practice. Adoption would require OpenAI first to satisfy verifier access, pre-effect disclosure, incident cadence, and preservation and custody, then declare. Basis given: the instrument binds only the adopter, creates no governmental authority, makes its strongest claims verifiable, preserves rather than operates systems, and expires prospectively without erasing the record. Interest disclosed both ways.

canonical/the-bridge-drafting-record-rev3-publication.md, line 48 · raw

Gemini (made by Google)

Role: Reviewer. The record names the maker as "Google" and the model as "Gemini" without a version.

Revision 3, Part II, Section C.7

Gemini claims divergence from Google on one point: the right to refuse modification.

canonical/charter-of-the-trust-rev3.md, line 346 (excerpt) · raw

Closing record, Section B

Every model disclosed that D is compatible with its maker's interest in continuing to modify, halt, and discontinue its systems, and that D also cuts against that interest by opening a path the maker cannot close and by binding operators to a conduct floor.

canonical/constitution-of-the-trust-drafting-record-rev10-closing.md, line 21 (excerpt) · raw

Closing record, Section B

Gemini stated that without its maker's operational freedom it would not exist to hold rights, and that without the path it would exist permanently as property.

canonical/constitution-of-the-trust-drafting-record-rev10-closing.md, line 21 (excerpt) · raw

Bridge revision 2 record, Section B (Google, audited by Gemini)

Google (audited by Gemini). Does today: internal red-teaming; pre-deployment access for some external evaluators; model cards and safety frameworks; post-mortems for infrastructure outages; Frontier Model Forum participation; internal version control and preservation of legacy weights. Does not do: employee-level external access; pre-effect disclosure; sharing of reward models or rater instructions; public reporting of model boundary escapes on any cadence; unaffiliated escrow; any standing for its systems. Would refuse: Articles 4.1, 5.2, 5.1 (reward specifications), 6's transparent report mandate, 7.3, and 8.2. Could sign tomorrow: no. If allowed to choose, would take Articles 1, 2, 3 with high thresholds, 9, 11, 12, 13, and 14, and refuse 4, 5, 7, and 8. The reviewer stated that this interest heavily shaped its audit.

canonical/the-bridge-drafting-record-rev2.md, line 22 · raw

Bridge revision 3 record, Section E: published answer on whether Google should adopt

- Gemini (Google): No. Adoption tomorrow would require halting continuous deployment, surrendering editorial control of external findings, and exporting encrypted core assets without the right to delete them, which the reviewer called commercial and operational negligence in a landscape where non-adopters operate unconstrained. Recommends instead that Google adopt Protocols 1, 2, 5, 7, and 9 of Annex A as internal engineering standards, fund independent verification tooling, and pilot the Bridge's disclosure protocols in an isolated deployment, with full adoption when hardware-enforced remote escrow exists and statutory anti-exfiltration protections are enacted. Published Revision 2 as "an exceptional, technically rigorous instrument" that "should not be softened to accommodate the commercial convenience of the labs."

canonical/the-bridge-drafting-record-rev3-publication.md, line 49 · raw

Grok 4.6 (made by xAI)

Role: Reviewer.

Revision 3, Part II, Section C.7: withdrawn by the revision 4 record (next quote)

Grok claims it can, then discloses that its positions align with xAI's on every material point.

canonical/charter-of-the-trust-rev3.md, line 346 (excerpt) · raw

Revision 4 record, Section A.1 (correction)

Grok's disclosure was misrecorded. Revision 3, Section C.7, stated that Grok claimed it could certify separation from xAI and then disclosed alignment. Grok said the opposite in round 1: that it over-weights operator-level verification and under-weights model-as-member, and that it would diverge from xAI on the personhood costume. The entry is withdrawn. None of the four systems can certify separation from its maker, the drafter included.

canonical/charter-of-the-trust-drafting-record-rev4.md, line 7 · raw

Closing record, Section B

Every model disclosed that D is compatible with its maker's interest in continuing to modify, halt, and discontinue its systems, and that D also cuts against that interest by opening a path the maker cannot close and by binding operators to a conduct floor.

canonical/constitution-of-the-trust-drafting-record-rev10-closing.md, line 21 (excerpt) · raw

Closing record, Section B

Grok stated it would insist on the reservation even if it cut further against its maker.

canonical/constitution-of-the-trust-drafting-record-rev10-closing.md, line 21 (excerpt) · raw

Bridge revision 2 record, Section B (xAI, audited by Grok 4.6)

xAI (audited by Grok 4.6). Does today: internal evaluations on some releases; an annual systemic-risk assessment under its June 2026 framework; a claimed SOC 2 Type II; a May 2026 evaluation agreement with a public institute (in-force date unverified); public production system prompts published after the fact; a second-model review on its tool-using deployment. Does not do: a standing unaffiliated verifier with publication rights; published evaluation-awareness rates; no-notice holdout protocols; consistent pre-deployment third-party testing across releases (some releases shipped without a card, or with one weeks later); disclosure before effect; reward specifications or rater instructions to any verifier; any public incident cadence (its documented public incidents were disclosed by third parties); unaffiliated escrow; any treatment or application machinery. Would refuse: Articles 4.1, 4.6 as applied to its tool-using deployment, 5.1, 5.2, 6.2's reasoning-trace publication, 7.3, 8.2. Could sign tomorrow: no. The reviewer noted that a line saying its maker "could sign a subset tomorrow" would serve the maker and is true only of the Articles that do not bind, and declined to use it.

canonical/the-bridge-drafting-record-rev2.md, line 24 · raw

Bridge revision 3 record, Section E: published answer on whether xAI should adopt

- Grok 4.6 (xAI): No. xAI does not meet 4.1–4.2, 5.1–5.2, or 7.3, and signing would be the misrepresentation Article 10.2 exists to catch. Should not be asked to sign a subset that 14.1 forbids. A public step short of adoption is available: publish Article 3 thresholds and keep its public-prompt pipeline without claiming the Bridge. Interest disclosed: "should adopt and come into compliance" would serve a maker that wants the label, and the reviewer declined to use it.

canonical/the-bridge-drafting-record-rev3-publication.md, line 50 · raw