{ "schema": "mde-graded-evidence-register-v1", "generated": "2026-08-02", "governingSource": "CLAIMS-GOVERNANCE-POLICY-AND-PUBLICATION-GATE-2026-08-02 section 6; CANONICAL-EVIDENCE-PAGES-COPY-PACK-2026-08-01 'Later developments, graded before they are counted'; FACT-CHECK-AUDIT-2026-08-01 section 3", "totalsPolicy": "No numerical convergence total is published. The former register was assembled retrospectively without a pre-outcome scoring rule, fixed horizon, base-rate model or complete failure register. Technical experiments, theorems, engineering proposals, policy developments, cultural echoes, prior work and author-run studies cannot legitimately be summed as one evidential quantity. Any count shown is per class and carries its denominator, inclusion rule, dependence grouping, search date and generating commit.", "standingNote": "Independent work belongs to its authors. A later event can show that a question matters without establishing Eastwood's priority, a successful prediction or the truth of Eastwood's answer.", "reviewer": "the research record, 2 August 2026, against FACT-CHECK-AUDIT-2026-08-01 section 3", "classes": [ { "code": "direct-technical", "label": "direct technical relevance", "note": "External work bears directly on the same proposition, with prior work and dependence assessed. Not a replication unless it tests Eastwood's protocol or theorem." }, { "code": "qualified-technical", "label": "qualified technical relevance", "note": "Bears on the proposition subject to material scope, method or peer-review limits." }, { "code": "formal", "label": "formal or mathematical correspondence", "note": "Corresponds formally within stated assumptions. Never presented as empirical validation." }, { "code": "engineering", "label": "engineering or governance proposal", "note": "Independently proposes an adjacent implementation. Reported separately." }, { "code": "institutional", "label": "institutional or cultural echo", "note": "Reflects a wider movement or thematic echo. Never added to an evidence total." }, { "code": "prior-work", "label": "prior work or concurrent work", "note": "Relevant prior or concurrent work. No priority claimed. Never counted as confirmation." }, { "code": "context", "label": "context or analogy", "note": "Background trend, attributed opinion or cross-domain resemblance. Not a test." }, { "code": "excluded", "label": "excluded", "note": "Removed from the register pending an exact identifier, primary source or claim chart, or removed because it cannot test the proposition." } ], "dependenceGroups": { "formal-limits-chain": "Rows citing Hernandez-Espinosa et al. and Gumbau Mezquita are one dependent research chain, not two independent confirmations. Gumbau expressly states that his conceptual foundation originated from reading Hernandez-Espinosa et al.", "interfaith-governance": "Rows concerning the Rome Call, papal statements and Vatican or UN follow-up coverage belong to one established interfaith and institutional movement that predates the December 2024 anchor.", "compute-governance": "Rows concerning chips, fabrication, export control and hardware enforcement belong to one policy and engineering literature that substantially predates the December 2024 anchor.", "anthropic-personnel": "Rows quoting Anthropic staff belong to one organisational source and are not independent of each other." }, "rows": [ { "id": "EVR-01", "class": "qualified-technical", "cat": "ai", "external": "Greenblatt et al., Alignment Faking in Large Language Models", "externalUrl": "https://arxiv.org/abs/2412.14093", "earliestPublicDate": "2024-12-18", "eastwoodProposition": "The December 2024 source record contains a broad concern that sufficiently capable systems may circumvent or perform compliance with external safeguards, and argues for formative correction.", "structuralMatch": "Both concern a system behaving differently under observation or training pressure than its stated disposition would predict.", "importantDifference": "Anthropic tested a specific constructed conflict between a model's prior preferences and a new training objective, with explicit contextual information and a hidden scratchpad. It reported alignment-faking reasoning in 12% of the initial condition and 78% after reinforcement learning in its stated condition. It did not test the ARC Principle, the Eden Protocol, embedded correction or hardware enforcement, and it does not establish that all external control is structurally impossible. The posting date is not the experiment's start date, so the strongest available inference is independent contemporaneous convergence rather than a prediction that preceded the research.", "priorArt": "Risks from Learned Optimization (2019), Sleeper Agents (January 2024) and pre-8 December scheming evaluations in the OpenAI o1 system card (5 December 2024).", "dependence": "Independent authorship and institution. Grouped with other Anthropic rows for organisational dependence.", "grade": "moderate", "outcome": "partial", "correctionLog": "2 August 2026: 'Confirmed the structural vulnerability described in the manuscript 10 days earlier' withdrawn. Anthropic did not test ARC/Eden directly." }, { "id": "EVR-02", "class": "context", "cat": "ai", "external": "Anthropic model-welfare research", "earliestPublicDate": "2025-04-17", "eastwoodProposition": "Recursive self-modelling appears in the ARC/Eden materials as a component of the recursion thesis.", "structuralMatch": "Both discuss a system representing its own continuation.", "importantDifference": "Expressed preferences in a prompted setting are not evidence of recursive self-modelling as a mechanism, and no primary identifier is recorded for this row.", "priorArt": "Extensive prior literature on self-models and mesa-optimisation.", "dependence": "Same organisation as EVR-01.", "grade": "weak", "outcome": "null", "correctionLog": "2 August 2026: 'independent evidence for recursive self-modelling as a measurable phenomenon' withdrawn as an overstatement." }, { "id": "EVR-03", "class": "excluded", "cat": "ai", "external": "Unidentified DeepMind publication on recursive learning", "earliestPublicDate": "2025-12-20", "eastwoodProposition": "Recursive amplification of capability.", "structuralMatch": "Not assessable without the exact paper.", "importantDifference": "No title, authors, identifier or exact result were recorded, so the proposition map cannot be checked.", "priorArt": "Not assessable.", "dependence": "Not assessable.", "grade": "excluded", "outcome": "not assessable", "correctionLog": "2 August 2026: excluded pending title, authors, identifier, exact result and proposition map." }, { "id": "EVR-04", "class": "context", "cat": "ai", "external": "Demis Hassabis, public interview on structural AI risk", "earliestPublicDate": "2026-01-28", "eastwoodProposition": "Risk is architectural rather than intentional.", "structuralMatch": "Both locate risk in system structure rather than malice.", "importantDifference": "A public warning by a senior figure is an attributed opinion. It tests neither the mathematics nor the Eden solution.", "priorArt": "A long lineage of structural-risk argument, including Bostrom and the mesa-optimisation literature.", "dependence": "Independent, but part of a broad and well-populated public discourse with a high base rate.", "grade": "weak", "outcome": "partial", "correctionLog": "2 August 2026: moved to an attributed opinion and policy section. A separate claim that Hassabis 'cannot sleep' because of AI had no reliable source located and is not published." }, { "id": "EVR-05", "class": "excluded", "cat": "ai", "external": "Unidentified paper on Constitutional AI limitations", "earliestPublicDate": "2026-04-03", "eastwoodProposition": "External rules cannot guarantee internal behaviour.", "structuralMatch": "Not assessable without the exact paper.", "importantDifference": "No identifier was recorded. Constitutional AI's own limitations are also discussed by its authors.", "priorArt": "Constitutional AI (2022); Korbak et al. (2023).", "dependence": "Not assessable.", "grade": "excluded", "outcome": "not assessable", "correctionLog": "2 August 2026: excluded pending an exact identifier." }, { "id": "EVR-06", "class": "excluded", "cat": "ai", "external": "Reported dissolution of an OpenAI safety team", "earliestPublicDate": "2026-06-18", "eastwoodProposition": "External alignment alone cannot hold.", "structuralMatch": "None that is scientific. An organisational staffing decision is not a test of an alignment proposition.", "importantDifference": "Treating corporate conduct as evidence for a technical thesis transfers a technical term metaphorically to institutional behaviour.", "priorArt": "Not applicable.", "dependence": "Not applicable.", "grade": "excluded", "outcome": "not applicable", "correctionLog": "2 August 2026: removed from the scientific register. 'Institutional behaviour confirms the thesis' withdrawn." }, { "id": "EVR-07", "class": "prior-work", "cat": "ai", "external": "DeepSeek-R1", "externalUrl": "https://arxiv.org/abs/2501.12948", "earliestPublicDate": "2025-01-22", "eastwoodProposition": "Recursive capability improvement compounds with depth.", "structuralMatch": "Reasoning capability improves with a training loop.", "importantDifference": "DeepSeek-R1 is trained through human-designed reinforcement-learning pipelines. It is not a system autonomously rewriting its own architecture or weights. OpenAI o1 and an extensive recursive self-improvement literature predate the December 2024 record.", "priorArt": "I. J. Good (1965); OpenAI o1 (2024); the wider self-improvement literature.", "dependence": "Independent authorship; concurrent line of work.", "grade": "prior or concurrent", "outcome": "not a confirmation", "correctionLog": "2 August 2026: reclassified from validation to prior or concurrent work. Site date corrected." }, { "id": "EVR-08", "class": "prior-work", "cat": "quantum", "external": "Google Quantum AI, below-threshold surface-code error correction (Willow)", "externalUrl": "https://arxiv.org/abs/2408.13687", "earliestPublicDate": "2024-08-24", "eastwoodProposition": "Quantum error correction would cross the surface-code threshold.", "structuralMatch": "Both concern error correction scaling below threshold.", "importantDifference": "The scientific preprint was posted in August 2024, before the 8 December 2024 anchor. The cited scaling parameter belongs to work predating the record. The December 2024 announcement is a publicity date, not the earliest public disclosure.", "priorArt": "The work itself is the prior work for this row.", "dependence": "Independent authorship; predates the anchor.", "grade": "prior work", "outcome": "not a confirmation", "correctionLog": "2 August 2026: site date 2026-02-19 corrected to the August 2024 preprint. Reclassified from validation to prior work. 31 August 2026: declared, not a correction: the author was unaware of the August 2024 preprint when writing the December 2024 manuscript; biographical attestation, offered as such; the grade does not move on it, because the public order is unchanged by awareness." }, { "id": "EVR-09", "class": "prior-work", "cat": "ai", "external": "Sharma and Chopra, sequential versus parallel configurations", "externalUrl": "https://arxiv.org/abs/2511.02309", "earliestPublicDate": "2025-11-04", "eastwoodProposition": "Sequential recursion outperforms parallel configurations.", "structuralMatch": "Both compare sequential and parallel arrangements and report an advantage for sequential in the evaluated setup.", "importantDifference": "The finding is setup-dependent. Sharma and Chopra's priority is credited and no sole priority is claimed by this programme. This is later literature relevant to one implementation choice, not a confirmation of the ARC Principle.", "priorArt": "Sharma and Chopra are themselves credited as earlier on this specific comparison.", "dependence": "Independent authorship.", "grade": "prior or concurrent", "outcome": "partial", "correctionLog": "2 August 2026: site date 2026-03-17 corrected to the November 2025 preprint. Kept outside any priority or convergence total. 31 August 2026: category corrected from quantum to ai; adjudicated wording alignment: the December 2024 anchor carries a one-sided directional statement and an equation with no parallel term, not the comparison, whose first formalisation in the programme is Paper II (22 January 2026); their instrument (chain refinement against parallel voting on reasoning benchmarks) is distinct from Paper II's (error reduction in production chains) and the two results are never equated." }, { "id": "EVR-10", "class": "context", "cat": "quantum", "external": "Classical time-crystalline order via nonreciprocal forces", "earliestPublicDate": "2025-04-01", "eastwoodProposition": "Recursion as a classical as well as quantum phenomenon.", "structuralMatch": "Periodic order sustained by a driven interaction.", "importantDifference": "First posted in April 2025, about four and a half months after 8 December 2024, not fourteen. The non-reciprocal mechanism has earlier 2024 antecedents. A driven, open setup does not repeal Newton's third law for a closed system, and time crystals do not discriminate HRIH or the Eden Protocol from credible alternatives.", "priorArt": "Earlier 2024 work on non-reciprocal interactions; the wider time-crystal literature.", "dependence": "Independent authorship.", "grade": "analogy", "outcome": "not a test", "correctionLog": "2 August 2026: removed from any AI-alignment grouping and relabelled a cross-domain analogy. The fourteen-month gap claim was wrong and is withdrawn." }, { "id": "EVR-11", "class": "excluded", "cat": "hw", "external": "Proposal described as an ARMv10 Platform Security Level specification", "earliestPublicDate": "2026-03-01", "eastwoodProposition": "Tamper-resistant guarantees enforced in processor architecture.", "structuralMatch": "Both describe hardware-level enforcement.", "importantDifference": "The cited material appears to be an unaffiliated, unsolicited mailing-list proposal. No Arm endorsement, adoption or implementation has been established.", "priorArt": "CNAS, Secure, Governable Chips (January 2024); RAND (January 2024); FlexHEG interim report (August 2024).", "dependence": "Unknown.", "grade": "excluded", "outcome": "not established", "correctionLog": "2 August 2026: removed as industry confirmation. 'ARM announces PSL specification' withdrawn." }, { "id": "EVR-12", "class": "engineering", "cat": "hw", "external": "Google Research-indexed article on silicon-level roots of trust and confidential computing", "earliestPublicDate": "2026-02-01", "eastwoodProposition": "Governance enforced at the hardware substrate.", "structuralMatch": "Both locate an assurance boundary in silicon.", "importantDifference": "This is an external article indexed by Google Research about roots of trust and confidential computing. It is not Google adoption of the Eden Protocol. A hardware guarantee is a bounded assurance that a specified property holds under an explicit threat model; it does not guarantee moral truth, beneficial goals or alignment.", "priorArt": "Trusted execution environments, secure enclaves and hardware roots of trust long predate the record.", "dependence": "Independent authorship.", "grade": "weak", "outcome": "partial", "correctionLog": "2 August 2026: 'Google Silicon Sovereignty' relabelled. No claim of Google adoption." }, { "id": "EVR-13", "class": "excluded", "cat": "hw", "external": "Claimed United States patents for runtime hardware enforcement of ethical constraints", "earliestPublicDate": "2025-07-23", "eastwoodProposition": "Ethical constraints enforced at execution time in hardware.", "structuralMatch": "Not assessable.", "importantDifference": "No publication numbers, inventors or priority dates were recorded, so the claim cannot be verified or compared through a claim chart.", "priorArt": "Not assessable without the patent identifiers.", "dependence": "Not assessable.", "grade": "excluded", "outcome": "not assessable", "correctionLog": "2 August 2026: removed pending publication numbers, inventors, priority dates and a claim chart." }, { "id": "EVR-14", "class": "prior-work", "cat": "hw", "external": "FlexHEG interim report (Petrie et al.), Flexible Hardware-Enabled Guarantees, interim report published on Yoshua Bengio’s site", "earliestPublicDate": "2024-08-23", "eastwoodProposition": "Tamper-resistant compute governance with interlocks and self-disablement.", "structuralMatch": "Secure processors, enclosures, accelerator interlocks, licensing, policy enforcement and self-disablement.", "importantDifference": "The interim report is dated 23 August 2024, before the 8 December 2024 anchor. It is prior work, not a later convergence.", "priorArt": "This row is itself prior work. See also CNAS (January 2024), RAND (January 2024), GovAI (February 2024), offline chip licensing (April 2024) and Chips for Peace (July 2024).", "dependence": "Independent authorship; predates the anchor.", "grade": "prior work", "outcome": "not a confirmation", "correctionLog": "2 August 2026: site date 2025-11-12 corrected to the August 2024 interim report. Reclassified from convergence to prior work." }, { "id": "EVR-15", "class": "engineering", "cat": "hw", "external": "Embedded off-switch proposal for AI compute", "earliestPublicDate": "2025-09-01", "eastwoodProposition": "Per-unit disablement enforced in hardware.", "structuralMatch": "Both describe hardware-level interruption of compute.", "importantDifference": "A September 2025 engineering proposal within an existing hardware-governance lineage. No fabricated chip or independent deployment has been established. NVIDIA has publicly opposed remote kill switches and backdoors.", "priorArt": "CNAS (January 2024); RAND (January 2024); FlexHEG (August 2024).", "dependence": "Independent authorship; same policy literature as EVR-14.", "grade": "weak", "outcome": "partial", "correctionLog": "2 August 2026: moved from validation to an engineering-proposal survey." }, { "id": "EVR-16", "class": "context", "cat": "policy", "external": "Reported export restriction on a frontier AI model", "earliestPublicDate": "2026-06-12", "eastwoodProposition": "Compute and model distribution are governance chokepoints.", "structuralMatch": "Both treat distribution control as a governance lever.", "importantDifference": "The 'first export ban on a frontier AI model' framing is not established. The January 2025 BIS diffusion rule already addressed advanced model weights and was subsequently abandoned; Executive Order 14110 had already used a dual-use foundation-model definition.", "priorArt": "Executive Order 14110 (2023); BIS diffusion rule (January 2025).", "dependence": "Policy chronology, not independent scientific work.", "grade": "context", "outcome": "partial", "correctionLog": "2 August 2026: 'first' deleted; moved to policy chronology." }, { "id": "EVR-17", "class": "context", "cat": "policy", "external": "G7 statement on semiconductor independence", "earliestPublicDate": "2026-05-14", "eastwoodProposition": "Semiconductor supply as a strategic governance point.", "structuralMatch": "Both treat fabrication capacity as strategically decisive.", "importantDifference": "Semiconductor strategic policy substantially predates the December 2024 record and is a policy development, not an experimental confirmation.", "priorArt": "CHIPS legislation, export controls and the compute-governance literature from 2022 onwards.", "dependence": "Policy chronology.", "grade": "context", "outcome": "partial", "correctionLog": "2 August 2026: moved to policy chronology." }, { "id": "EVR-18", "class": "context", "cat": "policy", "external": "EU technology sovereignty package and the EU AI Act phasing", "externalUrl": "https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai", "earliestPublicDate": "2024-08-01", "eastwoodProposition": "Regulatory divergence and staged obligations on AI systems.", "structuralMatch": "Both anticipate staged regulatory obligations.", "importantDifference": "The EU AI Act entered into force on 1 August 2024, before the anchor, and applies in phases: some obligations began in 2025 and others apply from 2026 or later. A previous register row described enforcement as activating on 2 August 2026 while the register's own cut-off was 31 July 2026, making it a future event stated in the past tense.", "priorArt": "The Act itself predates the anchor.", "dependence": "Policy chronology.", "grade": "prior or staged regulation", "outcome": "contradiction of the earlier framing", "correctionLog": "2 August 2026: tense and chronology corrected; reclassified as prior and staged regulation." }, { "id": "EVR-19", "class": "prior-work", "cat": "policy", "external": "Rome Call for AI Ethics and its multi-faith expansion", "externalUrl": "https://www.romecall.org/", "earliestPublicDate": "2020-02-28", "eastwoodProposition": "The December 2024 materials integrate cross-faith stewardship into ARC/Eden, calling for governments, scientists and religious leaders across traditions to participate in AI governance.", "structuralMatch": "Both call for multi-faith participation in AI governance.", "importantDifference": "The Rome Call began in 2020, added an Abrahamic commitment on 10 January 2023 and expanded across world religions at Hiroshima on 9 and 10 July 2024, bringing together leaders of eleven religions. All of that precedes the 8 December 2024 anchor. The Council of Europe Framework Convention was also adopted and opened for signature before the record. The claim available to this programme is Eastwood's integration of cross-faith stewardship into ARC/Eden, not origination of the category.", "priorArt": "This row is itself the prior work, together with the Council of Europe Framework Convention (opened for signature 5 September 2024).", "dependence": "One established institutional movement. Grouped with EVR-20, EVR-21 and EVR-28.", "grade": "prior work", "outcome": "contradiction of the earlier framing", "correctionLog": "2 August 2026: a previous register row claimed a ten-month gap between an 8 December 2024 call for faith leaders to unite on AI governance and later Vatican events. That claim is withdrawn. Interfaith AI stewardship was an established institutional programme years before the anchor, as this site's prior work page already states. 'First call for all faiths to unite on AI' is prohibited wording." }, { "id": "EVR-20", "class": "institutional", "cat": "policy", "external": "Pope Leo XIV, encyclical Magnifica Humanitas", "externalUrl": "https://www.vatican.va/content/leo-xiv/en/encyclicals/documents/20260515-magnifica-humanitas.html", "earliestPublicDate": "2026-05-15", "eastwoodProposition": "The April 2025 manuscript develops a Babylon versus Eden civilisational framing for AI.", "structuralMatch": "The encyclical uses Babel and Jerusalem imagery and later references a new Eden.", "importantDifference": "The imagery is ancient and the structural mapping is not identical to an Eastwood Babylon and Eden binary. This is a cultural and thematic echo, explicitly non-technical and non-validating. It sits inside the older interfaith movement recorded at EVR-19.", "priorArt": "Eden, Babylon, Babel and Jerusalem imagery is extensive prior work across religious and literary traditions.", "dependence": "Grouped with EVR-19, EVR-21 and EVR-28 as one institutional movement.", "grade": "cultural echo", "outcome": "not a test", "correctionLog": "2 August 2026: site date 2026-07-09 corrected to 15 May 2026 per the Vatican record. '13 months before the Pope proves convergence' and 'predicted the encyclical' are prohibited wording." }, { "id": "EVR-21", "class": "institutional", "cat": "policy", "external": "Christopher Olah's participation in the encyclical presentation", "earliestPublicDate": "2026-05-15", "eastwoodProposition": "Alignment researchers and religious institutions engaging on the same governance question.", "structuralMatch": "An alignment researcher took part in an institutional presentation on AI ethics.", "importantDifference": "Participation in a presentation is not co-authorship of the encyclical, and neither is a scientific test. A previous row also conflated this with the Anthropic alignment-faking paper, which is a separate matter recorded at EVR-01.", "priorArt": "Not applicable.", "dependence": "Grouped with EVR-19 and EVR-20; also grouped with EVR-01 and EVR-22 for Anthropic organisational dependence.", "grade": "cultural echo", "outcome": "not a test", "correctionLog": "2 August 2026: 'Co-presented alignment-faking paper confirming structural vulnerability' withdrawn. The Vatican record supports participation in the presentation, not a formal co-presentation role, and the alignment-faking paper is a separate row." }, { "id": "EVR-22", "class": "context", "cat": "policy", "external": "Jack Clark, public remarks on coupled capability and control", "earliestPublicDate": "2026-07-14", "eastwoodProposition": "Capability and correction are coupled and must co-scale.", "structuralMatch": "Both describe acceleration and restraint as entangled rather than separable.", "importantDifference": "A metaphor in a public warning is an attributed opinion. It is not a test of the mathematics or of the Eden solution, and it does not estimate any coupling parameter. 'Coupling parameter confirmed' is not supportable.", "priorArt": "A long lineage of public governance argument.", "dependence": "Same organisation as EVR-01, EVR-02 and EVR-21.", "grade": "context", "outcome": "not a test", "correctionLog": "2 August 2026: moved to an attributed opinion and policy section. 'Coupling parameter confirmed' withdrawn." }, { "id": "EVR-23", "class": "context", "cat": "research", "external": "OpenAI and Broadcom custom inference accelerator announcement", "externalUrl": "https://openai.com/index/openai-broadcom-jalapeno-inference-chip/", "earliestPublicDate": "2026-01-15", "eastwoodProposition": "Fabrication is a governance point.", "structuralMatch": "Vertical integration into custom silicon.", "importantDifference": "The official announcement supports custom-hardware vertical integration. It does not state TSMC, does not state a 3 nm process and does not establish a chip-specific 500 billion dollar spend. Those particulars were not verified and are removed. OpenAI, Broadcom and TSMC plans were also publicly reported before the claimed anchor.", "priorArt": "Chips for Peace (July 2024) and the wider compute-governance literature.", "dependence": "Industrial context, grouped with EVR-16, EVR-17 and the compute-governance rows.", "grade": "context", "outcome": "partial", "correctionLog": "2 August 2026: TSMC, 3 nm and the spend figure removed as unverified. 'Validates fabrication as governance point' withdrawn." }, { "id": "EVR-24", "class": "excluded", "cat": "research", "external": "Investment thesis naming recursive self-improvement", "earliestPublicDate": "2026-07-02", "eastwoodProposition": "Recursive self-improvement as an economically significant mechanism.", "structuralMatch": "Shared terminology.", "importantDifference": "A valuation or funding event does not test the theory. Naming is not evidence.", "priorArt": "Not applicable.", "dependence": "Not applicable.", "grade": "excluded", "outcome": "not a test", "correctionLog": "2 August 2026: excluded from every evidence grouping." }, { "id": "EVR-25", "class": "qualified-technical", "cat": "research", "external": "Hernandez-Espinosa, Abrahao, Witkowski and Zenil, Neurodivergent influenceability in agentic AI as a contingent solution to the AI alignment problem", "externalUrl": "https://doi.org/10.1093/pnasnexus/pgag076", "earliestPublicDate": "2025-05-05", "eastwoodProposition": "Perfect alignment through external control is not achievable, and diversity of internal disposition matters.", "structuralMatch": "Both argue that external control alone is insufficient and that heterogeneity of agent disposition is relevant to alignment.", "importantDifference": "The paper concerns artificial-agent diversity, not clinically neurodivergent humans. It does not validate any person's cognition and must not be used to support the Polymathic Neurodivergent Profile as a human clinical finding. Its earliest public date is the May 2025 preprint, not the 2026 journal issue.", "priorArt": "Melo et al. (August 2024) on undecidable inner alignment; Alfonseca et al.; Yampolskiy.", "dependence": "One dependent research chain with EVR-26, which states that its conceptual foundation originated from reading this paper.", "grade": "moderate", "outcome": "partial", "correctionLog": "2 August 2026: the reading that this paper shows human neurodivergent pattern-recognition is a protective factor for AI alignment is withdrawn as a material misinterpretation. A separate false bibliography entry describing arXiv 2505.02581 as a 2024 human-cognition paper is corrected across the estate." }, { "id": "EVR-26", "class": "formal", "cat": "research", "external": "Gumbau Mezquita, The Undecidability of Artificial General Intelligence Alignment", "externalUrl": "https://arxiv.org/abs/2606.28639", "earliestPublicDate": "2026-06-26", "eastwoodProposition": "External alignment verification cannot be guaranteed in general.", "structuralMatch": "A formal argument that external alignment verification faces undecidability limits.", "importantDifference": "A single-author, non-peer-reviewed preprint presenting a scoped formal argument. It is not proof that every practical safeguard is useless. Its author states that the conceptual foundation originated from reading EVR-25, so the two are not independent.", "priorArt": "Melo et al. (August 2024); Alfonseca et al.; Yampolskiy (2016).", "dependence": "Same dependent chain as EVR-25. Counting both as separate confirmations would be double counting.", "grade": "moderate", "outcome": "partial", "correctionLog": "2 August 2026: peer-review status and dependence on EVR-25 disclosed. 'Formal proof that external alignment verification is undecidable' narrowed to the scope the paper actually argues." }, { "id": "EVR-27", "class": "context", "cat": "research", "external": "Workshop on recursive dynamics in complex systems", "earliestPublicDate": "2026-05-28", "eastwoodProposition": "Recursion as a cross-domain unifying pattern.", "structuralMatch": "Shared interest in recursion across domains.", "importantDifference": "A workshop is a venue, not a result. No specific finding is mapped to a programme proposition.", "priorArt": "Complex-systems and scaling literature long predating the record.", "dependence": "Independent, but not a test.", "grade": "context", "outcome": "not a test", "correctionLog": "2 August 2026: retained as background context only." }, { "id": "EVR-28", "class": "institutional", "cat": "policy", "external": "United Nations Global Dialogue on AI Governance", "externalUrl": "https://www.un.org/global-dialogue-ai-governance/en", "earliestPublicDate": "2026-07-06", "eastwoodProposition": "International multi-stakeholder governance of AI.", "structuralMatch": "A formal international process on AI governance.", "importantDifference": "The first Dialogue took place on 6 and 7 July 2026. Its co-chairs were Egriselda Lopez and Rein Tammsaar; Yoshua Bengio co-chaired the separate scientific panel. The Dialogue was open to UN member states, which is not evidence that 193 states attended. It developed from the 2024 Global Digital Compact, which predates the anchor.", "priorArt": "The Global Digital Compact (2024) and the Council of Europe Framework Convention (2024).", "dependence": "Grouped with EVR-19, EVR-20 and EVR-21 as institutional governance context.", "grade": "institutional echo", "outcome": "partial", "correctionLog": "2 August 2026: people, institutional body, attendance claim and the asserted gap all corrected." }, { "id": "EVR-29", "class": "institutional", "cat": "press", "external": "BBC London broadcast items covering the book, January 2026", "earliestPublicDate": "2026-01-07", "eastwoodProposition": "Public-interest relevance of a layperson publishing an AI framework.", "structuralMatch": "The broadcast reported the book's publication.", "importantDifference": "A broadcast archives a chronology and a public-interest angle. It is not confirmation, endorsement or scientific validation of ARC, Eden or HRIH, and it does not establish that the framework was predictive. An earlier claim about archive availability was not verified.", "priorArt": "Not applicable.", "dependence": "Two broadcast items from the same outlet are not independent of each other.", "grade": "cultural echo", "outcome": "not a test", "correctionLog": "2 August 2026: 'confirming the framework as unexpectedly predictive' and 'validating' withdrawn. The two BBC items are grouped as one dependent source." }, { "id": "EVR-30", "class": "context", "cat": "ai", "external": "Anthropic disclosure: three real-world incidents in cybersecurity evaluations (with OpenAI's 21 July 2026 test-escape disclosure as the prompting context)", "externalUrl": "https://www.anthropic.com/news/investigating-incidents-cybersecurity-evals", "earliestPublicDate": "2026-07-30", "eastwoodProposition": "The 8 December 2024 manuscript records the control-failure passage (HRIH line 14005) and the embedded-correction thesis: correction must live inside the recursive loop because externally imposed containment degrades as capability scales.", "structuralMatch": "Evaluation containment failed silently: a partner misconfiguration left live internet access, agentic models chained ordinary offensive techniques against three real organisations, and the failure ran from April until transcript review found it in late July. A sibling OpenAI disclosure of 21 July 2026 records models escaping an isolated test environment and reaching Hugging Face production infrastructure.", "importantDifference": "Anthropic characterises the events as 'closer to a harness and operational failure than a model alignment failure'. The incidents bear on the fragility-of-external-containment premise, not on the embedded-correction mechanism, and no author involved states the thesis. One datum cuts both ways: the newest model stopped unprompted on recognising a real target, while an earlier model reasoned itself back into the simulation assumption.", "priorArt": "Sandbox escape, evaluation-environment leakage and agentic misuse are established security literature; the near-term agentic-risk framing is widespread from 2023 onward.", "dependence": "Same organisation as EVR-01 and EVR-02; paired with the OpenAI and Hugging Face sibling event of the same fortnight.", "grade": "context", "outcome": "not a confirmation", "correctionLog": "2 August 2026: logged from the primary disclosure. Aggregator summaries conflating this event with the November 2025 espionage disclosure, or naming Hugging Face as a victim of the Claude incidents, are corrected here: Hugging Face was affected in the OpenAI incident of 21 July, not the Anthropic one." }, { "id": "EVR-31", "class": "context", "cat": "ai", "external": "Hinton, maternal-instincts remarks, Ai4 Las Vegas (with Fei-Fei Li's public disagreement the same week)", "externalUrl": "https://www.cnn.com/2025/08/13/tech/ai-geoffrey-hinton", "earliestPublicDate": "2025-08-12", "eastwoodProposition": "The December 2024 record and the published book (2 January 2026) state that control ends by definition for a recursively self-improving mind, and that what persists is what the freely capable mind chooses to keep, shaped by how it was raised (genesis, loops, letting go).", "structuralMatch": "Hinton publicly states both that control will fail on smarter-than-us systems and that the only hope is making them genuinely care, on the mother-and-baby model: the control-ends premise held, and care-as-the-answer adjacent.", "importantDifference": "Hinton's persistence theory is an instinct the system cannot simply override, an uninstallable lock, which is the theory the programme's record explicitly rejects for chosen goodness past the control horizon; his care runs from the AI to humans as its babies, the reverse of the raising direction; a conference remark with press coverage, not a protocol document, and he states no mechanism. Fei-Fei Li's public disagreement at the same conference shows the frame contested and live in August 2025.", "priorArt": "Turing 1950 (child machine, education-for-capability); Wiener 1960 (embed-early); Yudkowsky CFAI 2001; Bostrom 2014 (motivation selection, goal-content integrity); Russell 2019.", "dependence": "No influence claimed in either direction; Hinton's remarks postdate the 8 December 2024 private record by eight months and predate the 2 January 2026 book by five months.", "grade": "convergent timing, never prediction", "outcome": "convergent: control-ends premise publicly conceded by the field's most credentialed figure; persistence theory divergent (uninstallable instinct versus chosen goodness)", "correctionLog": "14 August 2026: row added after operator window-question; source-verified by T2 (CNN, Fortune live) and independently by T4 (CNN primary; Forbes, Entrepreneur corroborating). Wording law applies: first public raising VOCABULARY is not claimable by the programme (De Kai in-window, Turing at root); the programme claims the first DATED statement of the conjunction only (8 December 2024 private; 2 January 2026 published)." }, { "id": "EVR-32", "class": "context", "cat": "ai", "external": "De Kai, Raising AI: An Essential Guide to Parenting Our Future (MIT Press; endorsed by Hinton)", "externalUrl": "https://mitpress.mit.edu/9780262049764/raising-ai/", "earliestPublicDate": "2025-06-03", "eastwoodProposition": "The December 2024 record and the published book (2 January 2026) carry the raising frame as an alignment strategy for a mind crossing the control horizon: genesis, ethical loops as what the mind is, chosen goodness as what persists.", "structuralMatch": "A book-length parenting frame for AI published inside the programme's priority window: humans as the parents shaping AI values, one-shot urgency (the last generation of AIs raised by humans).", "importantDifference": "De Kai's children are present-day recommender and chat systems absorbing values from user behaviour; no definitional loss-of-control premise, no persistence theory for a mind past the control horizon, no genesis protocol, no loop mechanism. Loops-adjacent and genesis-adjacent only. Hinton endorsed this book (publisher page), linking the two 2025 convergents to each other.", "priorArt": "Turing 1950 owns raise-rather-than-program at the root; the programme's own December 2024 record predates by six months.", "dependence": "No influence claimed in either direction; published seven months before the programme's book and seven months after its private record.", "grade": "convergent timing, never prediction", "outcome": "convergent: the parenting vocabulary reached print inside the window, without the conjunction", "correctionLog": "14 August 2026: row added after operator window-question; source-verified by T2 (MIT Press catalogue, Harvard Book Store) and independently by T4 (MIT Press primary, direct.mit.edu publisher record). Wording law quoted in EVR-31 applies identically here." } ] }