Consolidated account of retained research through 9 September 2026. This account accompanies r33 and two explicitly identified supplements. It is an interpretation of the acquired record, not a fresh audit of every original or a claim that the institutions below share one controller.
The investigation has moved beyond a list of people who know one another. We can identify recurring ways in which resources become influence: financing particular work, selecting the people who perform it, defining qualifying categories, supplying assessments, and attaching practical consequences to those assessments. An advertisement, a payment relationship, a platform account or access to research data can depend on decisions made across several institutions.
The most useful unit is the handoff. Who supplied an assessment? Under which agreement, rule or duty did someone else act on it? Who retained discretion? What actually happened? Who could reverse it? The record contains real financial and policy connections, along with shared methods and analogous structures. Those are different kinds of connection, and the coherent account depends on keeping them distinct.
Our strongest overall interpretation is that practical power is often distributed across these decision points. A central command is unnecessary for a restrictive result: independently motivated institutions can combine to produce it. That is an explanatory possibility supported by several bounded examples, not proof that every restriction arose this way. Common funders also support divergent positions, regulators decline proposals, distributors exit relationships, and courts or regulators impose remedies. Those observations constrain the stronger theory of a unified programme.
AI advocacy: capacity, selection and editorial authority
The original question remains important. The retained record shows substantial organised resources behind AI advocacy and several concrete routes into creator production. FLI's 2025 standalone audit reports US$45,197,838 in grants from Lightcone Foundation. Earlier transfers involving Lightcone Foundation and Future of Life Europe concern assets and organisational allocation; they cannot be added together as annual persuasion expenditure. Lightcone Foundation is also a different entity from Lightcone Infrastructure. Scale is real, but the accounting categories matter. A1
Creator influence is not one arrangement. People First's published terms provide for approval, revisions, compliance and conditional payment; an individual campaign statement of work was not acquired. Frame describes an independent creator track and an institutional-brief track, names funding sources, and denies funder control over content or selection. A Future Tense creator publicly acknowledged support from FLI, Frame and BlueDot while asserting editorial independence. Rational Animations and ControlAI acknowledged a partnership, but the fee and agreement remain unacquired. The sponsor behind the initial Sabine Hossenfelder account remains unidentified. A2
Selection can affect public communication without script approval. Supporting people who already share a concern can increase that concern's reach. That inference requires less than editorial command, but it still cannot identify an undisclosed payment or instruction. ControlAI's acquired UK articles locate formal management and membership authority in its directors. A small related-organisation payable in its US return does not name the counterparty and cannot complete an assumed UK–US funding chain. A3
The Global Call for AI Red Lines supplies a documented campaign-production chain: drafting, management, partner feedback and signatory work have named roles. CeSIA, The Future Society and CHAI are organisers; ControlAI and FLI provided partner feedback. The Call is distinct from ControlAI's A Narrow Path. Evan Hubinger's signature does not assign him every proposal advanced by another participant. Advocacy and summit access also remain separate from government adoption. This branch supplies real coordination, with a bounded object and purpose. A4
Shared resources do not settle the policy position
Oak's grants make an instructive comparison. Its published 2025 awards include three-year support to Check My Ads, Digital Freedom Fund and Statewatch. Statewatch endorsed objections to broad scanning and possible identity requirements in the EU child-abuse proposal; the WeProtect Secretariat supported mandatory detection with safeguards in a different, earlier version of that debate. The comparison is evidence of divergent interventions, not a same-draft controlled experiment.
DFF reports support for Liberty's challenge concerning the alleged Apple technical-capability notice. Liberty filed in March 2025, before the reported June 2026 grant. Funding could sustain existing litigation; it did not originate that filing. Acquired procedural results also do not determine the merits. A common donor is therefore an important financial connection, but not a sufficient explanation of every recipient's choices. A5
DFF's selection reforms further separate donor resources, project selection and award execution. Public descriptions assign project selection to external peers and information/facilitation roles to staff, with confidentiality and conflict commitments. They do not reveal every operative recusal, signature or payment decision. The question “who chooses?” needs an answer at each stage. A6
Policy travels through documents and decisions
The UK–Australia duty-of-care lane supplies stronger transmission evidence than a shared-event roster. Carnegie-associated work, collaboration with Reset and citations in Australia's statutory review connect a developed model to official consideration. Commissioned research and programme funding sustained parts of that policy infrastructure; later funding cannot be assigned authorship of an earlier idea.
The Australian exposure draft dated 8 September 2026 would allocate consequential discretion among Parliament's framework, ministerial rules and eSafety. It remains a proposal. Reset requested access for academic and third-sector researchers; the acquired draft instead requires employment by an Australian university and approval under prospective rules. That is a meaningful narrowing, not an awarded NGO entitlement. The private adult-consent provision limits what the duty requires; it does not create a universal adult veto. Nothing inspected establishes universal identification. A7
The UK record also distinguishes required outcomes from particular identity suppliers. Ofcom recognises different age-assurance methods and accepted email-based estimation after supplied evidence; that was a methodological choice, not vendor certification or procurement. Check My Ads later sought customer/identity checks for monetising creators and advertisers and pre-monetisation clearance where risk indicators appear. Those proposals could affect pseudonymity and income, but do not concern every ordinary viewer. Ofcom expressly considered its crisis-revenue request and did not recommend specific monetisation practices in that measure. Other recommendations retain their own unresolved dispositions. A8
Full Fact offers unusually precise evidence of input and institutional response. Ofcom attributes a notice-indicator amendment to its submission and questions from an earlier DSIT meeting. A notice becomes an indicator to consider; the provider retains its criteria-based crisis determination. The record supports a specific contribution while leaving the meeting's underlying records and other decisions unacquired. Influence is visible without assuming wholesale control. A8
The newly integrated procurement evidence makes the same distinction financially concrete. CDDP's Oliver Wyman evaluation commission has a disclosed £656,000 price excluding VAT, proposed criteria and a Treasury-linked feasibility gate. These establish commissioning and conditions, not payment, the final evaluation or an exercised funding decision. The announced £8.4 million programme allocation cannot be added to supplier prices as separate expenditure. Redactions and inconsistent contract dates remain unresolved. A23
OSDI had a CDEI-convened advisory group, PA Consulting programme evaluation and supplier participation by Faculty and PUBLIC. These are distinct roles. PUBLIC's account of a TRL4 sandbox describes prototype work; its later STCF2 assessment explicitly credits OSDI practices while retaining limited testing and supplier-reported performance. PUBLIC's claim to have contributed to CDDP does not identify it in a redacted subcontract list. Its current website-controller legal identity is not proof of the historical contracting party. The record now connects programmes, methods and personnel more clearly while preserving the missing agreements and acceptance decisions. A24
A classification matters because someone acts on it
The Lantern case study reports Meta sharing signals followed by separate investigations and action by Cash App and Xbox. The underlying case files were not inspected; the actions remain operator-reported. Programme rules include recourse, removal of no-longer-relevant information and administrator notification. Nothing in that case establishes that Lantern caused Steam removals. The live question is how information becomes a recipient decision, and what follows when the information changes. A9
Payment records make the recipient's role especially clear. Mastercard's retained August 2026 MATCH rules require inquiry but expressly permit onboarding a listed merchant. Stripe publishes a generally disqualifying practice with possible exceptions. The list and a restrictive underwriting decision are connected, but are not identical or a universal prohibition on banking. The 2025 eGumball appellate opinion applies FilmOn's reasoning about audience, purpose and context to confidential commercial reporting: an actual judicial bridge between advertising classification and payment reporting. Its procedural decision does not adjudicate the underlying report's truth or restore a merchant. A11
Other infrastructure gates allocate authority differently. Nominet's documented IWF cooperation predates the 2021 UK Safer Internet Centre grant, weakening the claim that the grant created that relationship. PIPCU combines public funding and police assessment with private intelligence; its announced industry co-funding pilot starts in April 2027 and does not establish current contributor command. Operation Creative's dated procedure allocates police tactical judgment, while advertising recipients choose how to implement information. Broader KYBC requests sought identification and consequences for commercial customers; requested reach across hosting and payments is not enacted universal user-ID law. A13
Advertising classification adds a question about scale: is the object a page, channel or whole domain? February 2025 Senate letters questioned financing a website even where advertisements avoided prohibited material. DoubleVerify subsequently announced default Highly Illicit and P2P Sharing and Streaming categories and later reported implementation. The chronology does not establish that the letters uniquely caused the change. NCMEC notifications, DV's assessment and its later IWF relationship are different inputs. Announced defaults do not constitute our observation of customer settings, and general customer-choice language does not prove an override in every configuration. A14
Counterpressure belongs in this account. TAG reported no violations of the investigated brand-safety guidelines: its determination, not independent exoneration. DV's complaint challenged critics' assertions: allegations, not judicial findings. Its disclosed fee model does not establish a bounty for blacklisting. The consequences can be significant without a proven private motive or a single upstream instruction.
GDI illustrates why time matters. Its advertising-risk data had a documented Oracle distribution route, but Oracle's relevant service ended in September 2024. IAS later said it discontinued limited past GDI inputs. Historical relationships cannot establish present delivery. A publisher-review procedure may revise GDI's assessment, but the acquired record does not show automatic propagation to every buyer or an obligation to resume advertising. A16
Research access has its own gatekeepers
GDI's public funding and research work reveal further roles without completing an alleged government blacklist. FCDO described supervised activities outside the United States; the signed grant and commercial-reuse conditions remain unacquired. CORDIS identifies a German GDI entity and a €385,275 EU contribution to AI4TRUST, not the whole consortium budget or an assumed UK guarantee. Project documents show keywords and annotations, model-assisted triage and human review. A seed term is not a verdict; a research warning is not an advertising ban. GDI's request for mandatory tool/API compatibility is a bid for an institutional role, not proof of its award. A17
The integrated EU access rules separate three mechanisms. Article 40(4) provides project-specific vetting and a reasoned request through the relevant Digital Services Coordinator. Article 40(12) concerns public-data access. Article 22 trusted-flagger priority concerns notices and is different again. Funding and independence disclosures can form part of eligibility; public overviews need not name every applicant or funder. Neither Horizon funding nor a failed exact-name search establishes whether GDI received platform access. A18
AliExpress's June 2025 binding commitments supply concrete procedures and clocks. The company nominates and pays a monitoring trustee; the Commission approves and supervises the role. Paying the trustee does not confer the Commission's approval power, and appointment does not establish delivery of a report. The substantive commitment term and external reporting schedule differ. A later illegal-product fine concerns a separate issue; it is not evidence of breach of the research-access commitments. A19
Assurance is not the same evidence as a delivered result
The retained FTI audit offers a useful stress test. It gives a positive Article 40(12) opinion while reporting no approved applications in the year ending June 2025. It says all examined applications were incomplete or invalid. The record supports a narrower conclusion about assessed readiness. It does not establish either that meaningful access was delivered or that requests were improperly refused. Its period also precedes the monitoring trustee's appointment. A20
Two explanations remain possible: legitimate screening of deficient requests, or an interpretation/cure process that makes an access right difficult to use. Without application files, responses and actual delivery, the public audit cannot decide between them. Likewise, the first monitoring report and the Commission's response were not acquired; absence from the acquired record is not proof of non-delivery.
The researcher-side material is informative but not an adjudication. The DSA40 Collaboratory collects voluntary reports, assists applicants and advocates support for intermediaries. Its September 2026 tracker headline reports 70 applications, while its retained table sums to 68; other subtotals also differ. Its one open AliExpress entry lacks the identity and dates needed to connect it to GDI or the FTI cohort. Organisers' judgments that rejection reasons are invalid are leads, not acquired rulings. Disclosed project and host funding establish resources, not donor control of findings. A21
This is the same evidentiary problem seen in the procurement branch: an evaluator commissioned, a test described, a positive opinion, an accepted application and a useful result are distinct events. Counting one as another can make both success and obstruction appear more certain than the record permits.
Can correction travel as far as the consequence?
The retained public ThreatExchange client provides a substantive counterexample to technical permanence. At the inspected commit, deletion markers can remove local records; a stale-checkpoint path calls for clearing and a fresh fetch; opinion structures can represent disagreement. These are static capabilities in particular code paths, not a demonstrated end-to-end deployed execution trace. Lantern's use of hosted ThreatExchange does not prove its participants use that client or version. No account restoration was observed. TCAP's separately inspected hash interface also documents deletion and feedback; a historical Altitude importer is another interface and version. A10
There are also actual obligations constraining recipient discretion. The January 2025 CFPB order requires Block to provide specified restriction/appeal notices and to decide an appeal and change status as appropriate within ten business days. That does not guarantee every appellant restored access or constitute an admission of every finding. It does materially narrow the earlier remedy gap. Acquired July settlement terms likewise require careful separation of assent, court entry, credited redress, reimbursement and actual payment. A12
The unresolved question is therefore sharper than “is there an appeal?” Does a correction reach the recipient? Does it reach exported copies and enforcement systems? Is a completed case reopened? Is access, money or reputation restored where warranted? The most revealing missing artifact would be an ordinary operational trace connecting those events—not an implausible confession of hidden intent.
The coherent picture—and its remaining breaks
The record now supports several connected institutional stories, with specific cross-domain bridges and consequential local powers. It also supports a comparative model: resources sustain capacity; selection and definitions shape what is assessed; assessments can condition access; recipients may add their own restrictions; and correction must cross the same institutional boundaries as the original action. These roles are a way of investigating the systems, not proof that all of them form one coordinated pipeline. A22
The original broad hypothesis of common direction remains open. Evidence of local coordination has strengthened considerably, while evidence of divergence and retained discretion makes a single-controller account less adequate as a default explanation. No acquired chain establishes the broader Fabian/WEF/owner-bank/medical-data coordination suggested earlier. Those remain hypotheses or unresolved leads, not conclusions smuggled in through visual proximity.
The dated stablecoin supplement reaches a similar boundary: BlackRock's reserve management, BNY's custody and Circle's issuance are different functions. The retained check examines actual control surfaces but establishes no instruction chain from them to Steam's restrictions. It remains a bounded 9 September check, with no live-key audit or new current-law verification. A15
The next most discriminating evidence lies between the roles already identified: a donor's actual terms and exercised approval; a redacted supplier's real agreement; a classification revision received and acted on by a customer; an application paired with its response and usable data; an assurance report paired with corrective action. That is where the investigation can turn a plausible pattern into a demonstrated mechanism—or rule one out.
Inspect the claim and evidence index. The full atlas retains detailed relationships that this account cannot reasonably narrate. Source counts describe records, not independent witnesses; several edges can rest on one document. This consolidation reconciles accepted material and selected stale questions. It does not claim every surviving lead has been exhausted.