Skip to content

AI Safety Observatory

Claims and provenance.

Developments, papers and corrections. Sources first; uncertainty kept in view.

Ordered by the date each record carries, oldest first. Private records are labelled as such; a private record is not a public disclosure.

Historical edition date: 2026-09-20 · Weekly updates planned

Counted separately: 19 convergences listed by Paper XI · 41 dated claims · 18 claim families · 22 registered propositions · 12 tracked predictions and 10 outcome rows · 33 external records · 10 antecedents conceded · 11 dated records · 3 open checks · 0 independent replication attempts recorded. These are different kinds of thing and are never added together.

Dated records, external work and antecedents, oldest firstDated recordExternal workAntecedent1 January 1965Good, I. J. (1965). Speculations Concerning the First Ultraintelligent MachineGood, I. J. (1965). Speculations Concern...30 March 1993Vinge, V. (1993). The Coming Technological SingularityVinge, V. (1993). The Coming Technologic...4 April 1997West, G. B., Brown, J. H., and Enquist, B. J. (1997). A General Model for the Origin of Allometric Scaling Laws in Biology. Science, 276(5309):122-126. doi:10.1126/science.276.5309.122West, G. B., Brown, J. H., and Enquist, ...1 April 2003White, C. R., and Seymour, R. S. (2003). Mammalian basal metabolic rate is proportional to body mass to the two-thirds power. Proceedings of the National Academy of Sciences, 100(7):4046-4049. doi:10.1073/pnas.0436428100White, C. R., and Seymour, R. S. (2003)....1 April 2004Kozlowski, J., and Konarzewski, M. (2004). Is West, Brown and Enquist's model of allometric scaling mathematically correct and biologically relevant? Functional Ecology, 18(2):283-289. doi:10.1111/j.0269-8463.2004.00830.xKozlowski, J., and Konarzewski, M. (2004...24 April 2007Bettencourt, L. M. A., Lobo, J., Helbing, D., Kuhnert, C., and West, G. B. (2007). Growth, innovation, scaling, and the pace of life in cities. Proceedings of the National Academy of Sciences, 104(17):7301-7306. doi:10.1073/pnas.0610172104Bettencourt, L. M. A., Lobo, J., Helbing...28 February 2012Hutter, M. (2012). Can Intelligence Explode? Journal of Consciousness Studies, 19(1-2):143-166. arXiv:1202.6177Hutter, M. (2012). Can Intelligence Expl...3 July 2014Bostrom, N. (2014). Superintelligence, chapters 8 to 13Bostrom, N. (2014). Superintelligence, c...28 February 2020Rome Call for AI Ethics and its multi-faith expansionRome Call for AI Ethics and its multi-fa...19 July 2020Yampolskiy, R. V. (2020). On Controllability of AI. arXiv:2008.04071Yampolskiy, R. V. (2020). On Controllabi...1 August 2024EU technology sovereignty package and the EU AI Act phasingEU technology sovereignty package and th...6 August 2024Snell, C., Lee, J., Xu, K. and Kumar, A. (2024). Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters. arXiv:2408.03314Snell, C., Lee, J., Xu, K. and Kumar, A....23 August 2024FlexHEG interim report (Petrie et al.), Flexible Hardware-Enabled Guarantees, interim report published on Yoshua Bengio’s siteFlexHEG interim report (Petrie et al.), ...24 August 2024Google Quantum AI, below-threshold surface-code error correction (Willow)Google Quantum AI, below-threshold surfa...8 December 20248 December 2024 manuscript8 December 2024 manuscript9 December 2024Paper XI: Willow: first below-threshold quantum error correctionPaper XI: Willow: first below-threshold ...18 December 2024Greenblatt et al., Alignment Faking in Large Language ModelsGreenblatt et al., Alignment Faking in L...22 January 2025DeepSeek-R1DeepSeek-R11 April 2025Classical time-crystalline order via nonreciprocal forcesClassical time-crystalline order via non...17 April 2025Anthropic model-welfare researchAnthropic model-welfare research30 April 202530 April 2025 manuscript30 April 2025 manuscript5 May 2025Hernandez-Espinosa, Abrahao, Witkowski and Zenil, Neurodivergent influenceability in agentic AI as a contingent solution to the AI alignment problemHernandez-Espinosa, Abrahao, Witkowski a...3 June 2025De Kai, Raising AI: An Essential Guide to Parenting Our Future (MIT Press; endorsed by Hinton)De Kai, Raising AI: An Essential Guide t...23 July 2025Claimed United States patents for runtime hardware enforcement of ethical constraintsClaimed United States patents for runtim...12 August 2025Hinton, maternal-instincts remarks, Ai4 Las Vegas (with Fei-Fei Li's public disagreement the same week)Hinton, maternal-instincts remarks, Ai4 ...1 September 2025Embedded off-switch proposal for AI computeEmbedded off-switch proposal for AI comp...9 September 2025Paper XI: Embedded off-switches for AI compute: deadman-switch security anchorsPaper XI: Embedded off-switches for AI c...4 November 2025Sharma and Chopra, sequential versus parallel configurationsSharma and Chopra, sequential versus par...4 November 2025Sharma and Chopra, sequential versus parallel configurationsSharma and Chopra, sequential versus par...20 December 2025Unidentified DeepMind publication on recursive learningUnidentified DeepMind publication on rec...2 January 2026Infinite Architects paperbackInfinite Architects paperback6 January 2026Infinite Architects ebookInfinite Architects ebook7 January 2026BBC London broadcast items covering the book, January 2026BBC London broadcast items covering the ...15 January 2026OpenAI and Broadcom custom inference accelerator announcementOpenAI and Broadcom custom inference acc...17 January 2026Paper IPaper I22 January 2026Paper IIPaper II28 January 2026Demis Hassabis, public interview on structural AI riskDemis Hassabis, public interview on stru...1 February 2026Google Research-indexed article on silicon-level roots of trust and confidential computingGoogle Research-indexed article on silic...6 February 2026Paper XI: Classical time crystal via nonreciprocal forcesPaper XI: Classical time crystal via non...9 February 2026Paper IIIPaper III13 February 2026Foundational paperFoundational paper22 February 2026Eden Engineering public noteEden Engineering public note1 March 2026Proposal described as an ARMv10 Platform Security Level specificationProposal described as an ARMv10 Platform...16 March 2026Paper IV seriesPaper IV series16 March 2026Papers V through IXPapers V through IX3 April 2026Unidentified paper on Constitutional AI limitationsUnidentified paper on Constitutional AI ...17 April 2026Paper XI: Recursive Superintelligence: $500M at $4B for closed-loop recursive self-improvement; convergent naming with A1's core conceptPaper XI: Recursive Superintelligence: $...14 May 2026G7 statement on semiconductor independenceG7 statement on semiconductor independen...15 May 2026Pope Leo XIV, encyclical Magnifica HumanitasPope Leo XIV, encyclical Magnifica Human...15 May 2026Christopher Olah's participation in the encyclical presentationChristopher Olah's participation in the ...25 May 2026Paper XI: Pope Leo XIV first encyclical: Babel versus Jerusalem framing for AIPaper XI: Pope Leo XIV first encyclical:...25 May 2026Paper XI: Christopher Olah (Anthropic co-founder) co-presents encyclical at the VaticanPaper XI: Christopher Olah (Anthropic co...28 May 2026Workshop on recursive dynamics in complex systemsWorkshop on recursive dynamics in comple...3 June 2026Paper XI: EU Tech Sovereignty Package (Chips Act 2.0 + CADA)Paper XI: EU Tech Sovereignty Package (C...12 June 2026Reported export restriction on a frontier AI modelReported export restriction on a frontie...12 June 2026Paper XI: US orders Fable 5 + Mythos 5 offline: first export-control action against a deployed frontier modelPaper XI: US orders Fable 5 + Mythos 5 o...18 June 2026Reported dissolution of an OpenAI safety teamReported dissolution of an OpenAI safety...19 June 2026Paper XI: Macron at G7 Evian: 'trusted partners' compute sovereignty framingPaper XI: Macron at G7 Evian: 'trusted p...19 June 2026Paper XI: Jack Clark (Anthropic co-founder): 'gas pedal but no brake pedal'Paper XI: Jack Clark (Anthropic co-found...26 June 2026Gumbau Mezquita, The Unverifiability of Artificial General Intelligence (AGI) Alignment, Static and Dynamic: From Trakhtenbrot's Wall to the Safety-Generality Tension (arXiv 2606.28639 v2, 6 July 2026; v1 of 26 June 2026 was titled The Undecidability of Artificial General Intelligence (AGI) Alignment)Gumbau Mezquita, The Unverifiability of ...2 July 2026Investment thesis naming recursive self-improvementInvestment thesis naming recursive self-...6 July 2026United Nations Global Dialogue on AI GovernanceUnited Nations Global Dialogue on AI Gov...14 July 2026Jack Clark, public remarks on coupled capability and controlJack Clark, public remarks on coupled ca...30 July 2026Anthropic disclosure: three real-world incidents in cybersecurity evaluations (with OpenAI's 21 July 2026 test-escape disclosure as the prompting context)Anthropic disclosure: three real-world i...
Every dated object in the provenance record on one axis: the programme's dated records (squares, each labelled private or public by its own basis), external work by its earliest public date (circles) and conceded antecedents (triangles). Position on the axis is chronology only, never evidential support. Drawn from the registers named on this page.
Table view
DateRecordKindDate basis
Good, I. J. (1965). Speculations Concerning the First Ultraintelligent MachineAntecedentantecedent
Vinge, V. (1993). The Coming Technological SingularityAntecedentantecedent
West, G. B., Brown, J. H., and Enquist, B. J. (1997). A General Model for the Origin of Allometric Scaling Laws in Biology. Science, 276(5309):122-126. doi:10.1126/science.276.5309.122Antecedentantecedent
White, C. R., and Seymour, R. S. (2003). Mammalian basal metabolic rate is proportional to body mass to the two-thirds power. Proceedings of the National Academy of Sciences, 100(7):4046-4049. doi:10.1073/pnas.0436428100Antecedentantecedent
Kozlowski, J., and Konarzewski, M. (2004). Is West, Brown and Enquist's model of allometric scaling mathematically correct and biologically relevant? Functional Ecology, 18(2):283-289. doi:10.1111/j.0269-8463.2004.00830.xAntecedentantecedent
Bettencourt, L. M. A., Lobo, J., Helbing, D., Kuhnert, C., and West, G. B. (2007). Growth, innovation, scaling, and the pace of life in cities. Proceedings of the National Academy of Sciences, 104(17):7301-7306. doi:10.1073/pnas.0610172104Antecedentantecedent
Hutter, M. (2012). Can Intelligence Explode? Journal of Consciousness Studies, 19(1-2):143-166. arXiv:1202.6177Antecedentantecedent
Bostrom, N. (2014). Superintelligence, chapters 8 to 13Antecedentantecedent
Rome Call for AI Ethics and its multi-faith expansionExternal workearliest public
Yampolskiy, R. V. (2020). On Controllability of AI. arXiv:2008.04071Antecedentantecedent
EU technology sovereignty package and the EU AI Act phasingExternal workearliest public
Snell, C., Lee, J., Xu, K. and Kumar, A. (2024). Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters. arXiv:2408.03314Antecedentantecedent
FlexHEG interim report (Petrie et al.), Flexible Hardware-Enabled Guarantees, interim report published on Yoshua Bengio’s siteExternal workearliest public
Google Quantum AI, below-threshold surface-code error correction (Willow)External workearliest public
8 December 2024 manuscriptDated recordprivate-record
Paper XI: Willow: first below-threshold quantum error correctionExternal workexternal date, Paper XI
Greenblatt et al., Alignment Faking in Large Language ModelsExternal workearliest public
DeepSeek-R1External workearliest public
Classical time-crystalline order via nonreciprocal forcesExternal workearliest public
Anthropic model-welfare researchExternal workearliest public
30 April 2025 manuscriptDated recordprivate-record
Hernandez-Espinosa, Abrahao, Witkowski and Zenil, Neurodivergent influenceability in agentic AI as a contingent solution to the AI alignment problemExternal workearliest public
De Kai, Raising AI: An Essential Guide to Parenting Our Future (MIT Press; endorsed by Hinton)External workearliest public
Claimed United States patents for runtime hardware enforcement of ethical constraintsExternal workearliest public
Hinton, maternal-instincts remarks, Ai4 Las Vegas (with Fei-Fei Li's public disagreement the same week)External workearliest public
Embedded off-switch proposal for AI computeExternal workearliest public
Paper XI: Embedded off-switches for AI compute: deadman-switch security anchorsExternal workexternal date, Paper XI
Sharma and Chopra, sequential versus parallel configurationsExternal workearliest public
Sharma and Chopra, sequential versus parallel configurationsExternal workearliest public
Unidentified DeepMind publication on recursive learningExternal workearliest public
Infinite Architects paperbackDated recordpublic-disclosure
Infinite Architects ebookDated recordpublic-disclosure
BBC London broadcast items covering the book, January 2026External workearliest public
OpenAI and Broadcom custom inference accelerator announcementExternal workearliest public
Paper IDated recordpublic-disclosure
Paper IIDated recordunstated
Demis Hassabis, public interview on structural AI riskExternal workearliest public
Google Research-indexed article on silicon-level roots of trust and confidential computingExternal workearliest public
Paper XI: Classical time crystal via nonreciprocal forcesExternal workexternal date, Paper XI
Paper IIIDated recordunstated
Foundational paperDated recordunstated
Eden Engineering public noteDated recordunstated
Proposal described as an ARMv10 Platform Security Level specificationExternal workearliest public
Paper IV seriesDated recordunstated
Papers V through IXDated recordunstated
Unidentified paper on Constitutional AI limitationsExternal workearliest public
Paper XI: Recursive Superintelligence: $500M at $4B for closed-loop recursive self-improvement; convergent naming with A1's core conceptExternal workexternal date, Paper XI
G7 statement on semiconductor independenceExternal workearliest public
Pope Leo XIV, encyclical Magnifica HumanitasExternal workearliest public
Christopher Olah's participation in the encyclical presentationExternal workearliest public
Paper XI: Pope Leo XIV first encyclical: Babel versus Jerusalem framing for AIExternal workexternal date, Paper XI
Paper XI: Christopher Olah (Anthropic co-founder) co-presents encyclical at the VaticanExternal workexternal date, Paper XI
Workshop on recursive dynamics in complex systemsExternal workearliest public
Paper XI: EU Tech Sovereignty Package (Chips Act 2.0 + CADA)External workexternal date, Paper XI
Reported export restriction on a frontier AI modelExternal workearliest public
Paper XI: US orders Fable 5 + Mythos 5 offline: first export-control action against a deployed frontier modelExternal workexternal date, Paper XI
Reported dissolution of an OpenAI safety teamExternal workearliest public
Paper XI: Macron at G7 Evian: 'trusted partners' compute sovereignty framingExternal workexternal date, Paper XI
Paper XI: Jack Clark (Anthropic co-founder): 'gas pedal but no brake pedal'External workexternal date, Paper XI
Gumbau Mezquita, The Unverifiability of Artificial General Intelligence (AGI) Alignment, Static and Dynamic: From Trakhtenbrot's Wall to the Safety-Generality Tension (arXiv 2606.28639 v2, 6 July 2026; v1 of 26 June 2026 was titled The Undecidability of Artificial General Intelligence (AGI) Alignment)External workearliest public
Investment thesis naming recursive self-improvementExternal workearliest public
United Nations Global Dialogue on AI GovernanceExternal workearliest public
Jack Clark, public remarks on coupled capability and controlExternal workearliest public
Anthropic disclosure: three real-world incidents in cybersecurity evaluations (with OpenAI's 21 July 2026 test-escape disclosure as the prompting context)External workearliest public

The dated records

Each record is listed with the basis on which its date was fixed, in the register's own words. A self-sent email is a private record; the public disclosure dates are the book editions and the paper releases.

EP-01 · · private record: a self-sent email, dated by the mail server; not a public disclosure

8 December 2024 manuscript

The future of intelligence may not only surpass biological constraints but also transcend traditional definitions of what it means to be “alive”

8 December 2024 manuscript, 8 December 2024, line 429 · SHA-256 ae3bffe13821 · authored
Artefact
Manuscript essay establishing the embedded-correction alignment thesis
Artefact type
Essay manuscript (Version 2), self-emailed bundle of five attachments
How the date was fixed
Date and Message-ID assigned by the sender's mail servers on send; existence-by-date anchor.
What it establishes
The embedded-correction alignment thesis: alignment must be embedded in the substrate of an AI system rather than imposed from outside; the ARC Principle U = I × R (utility scales as the product of intelligence and recursion) is named at lines 6 and 9, with the equation first appearing at line 495; the AI control-failure prediction ('AI systems cannot be truly controlled. They will evolve beyond any safeguards') appears at HRIH line 14005, ten days before Anthropic's alignment-faking paper (Greenblatt et al., arXiv:2412.14093, 18 December 2024). The bundle also carries the name 'Eden Protocol' itself, in the book draft v3.2 (Chapter 20), anchoring the name to this date.
SHA-256
f0d1f38f
Primary source
https://www.michaeldariuseastwood.com/research/evidence-spine.html
Claims first dated here
PC-001, PC-002, PC-003, PC-004, PC-005, PC-007
Match basis
claim first_date equals the episode date
Quotations on record for this artefact
255 distinct (one shown, the highest scoring)

EP-02 · · private record: a self-sent email, dated by the mail server; not a public disclosure

30 April 2025 manuscript

Conclusion: A World and Possibly a Cosmos Bound by Empathy Here, near the end of Chapter Eight, I want to underline that the best governance for AI is not an afterthought but a living tapestry that merges moral doping, meltdown trig- gers, orchard caretaker metaphors, and universal collabora- tion.

30 April 2025 manuscript, 30 April 2025, page 334 · SHA-256 735a975cca2c · authored
Artefact
Manuscript package expanding the Eden Protocol and first naming Embedded Correction and Meltdown Triggers
Artefact type
Manuscript package,.eml self-email
How the date was fixed
Gmail server timestamp on self-sent email
What it establishes
Expanded articulation of the Eden Protocol for embedded / hardware-level AI alignment, developing the 8 December 2024 record, which already carries the name; first named use of Embedded Correction and Meltdown Triggers.
Primary source
https://www.michaeldariuseastwood.com/research/evidence-spine.html
Claims first dated here
PC-009
Match basis
claim first_date equals the episode date
Quotations on record for this artefact
329 distinct (one shown, the highest scoring)

EP-03 · · public disclosure date

Infinite Architects paperback

Artefact
Book: Infinite Architects: Intelligence, Recursion, and the Creation of Everything (paperback)
Artefact type
Book (print edition)
How the date was fixed
Public print publication date
What it establishes
First public book-length articulation of the embedded-correction alignment thesis, ARC Principle, Eden Protocol, and 37 named concepts.
Primary source
https://www.michaeldariuseastwood.com/#book
Claims first dated here
PC-010, PC-011, PC-012, PC-013, PC-014, PC-015, PC-016, PC-017, PC-018, PC-019, PC-020, PC-021
Match basis
claim first_date equals the episode date
Quotations on record for this artefact
none

EP-04 · · public disclosure date

Infinite Architects ebook

This is exactly what we seek with meltdown alignment: a state where the values are so deeply embedded that deviation becomes impossible not because it is forbidden but because the system has become oriented around those values.

Infinite Architects ebook, 6 January 2026 · SHA-256 9d10d64861b7 · authored
Artefact
Book: Infinite Architects (ebook edition)
Artefact type
Book (ebook edition)
How the date was fixed
Public ebook publication date
What it establishes
Public ebook publication of Infinite Architects, four days after the print edition.
Primary source
https://www.amazon.co.uk/Infinite-Architects-Intelligence-Recursion-Everything-ebook/dp/B0GDQZ7P21
Claims first dated here
PC-022
Match basis
claim first_date equals the episode date
Quotations on record for this artefact
175 distinct (one shown, the highest scoring)

EP-05 · · public disclosure date

Paper I

Artefact
The ARC Principle: Formalisation and Preliminary Validation of Recursive Capability Scaling (Paper I)
How the date was fixed
Public paper release
What it establishes
First public formal presentation of the ARC Principle, with preliminary validation of super-linear error suppression under sequential recursive processing.
DOI
10.17605/OSF.IO/6C5XB
Primary source
https://www.michaeldariuseastwood.com/research/papers/paper-i-arc-principle.html
Claims first dated here
PC-024, PC-040
Match basis
claim first_date equals the episode date
Quotations on record for this artefact
none

EP-06 · · date basis not stated in the register

Paper II

Artefact
The ARC Principle: Experimental Validation Across Frontier Models (Paper II)
How the date was fixed
not stated in the register
What it establishes
Experimental validation of the ARC Principle across multiple frontier models.
DOI
10.17605/OSF.IO/6C5XB
Primary source
https://www.michaeldariuseastwood.com/research/papers/paper-ii-experimental-validation.html
Claims first dated here
PC-025, PC-026, PC-027, PC-028, PC-029, PC-030
Match basis
claim first_date equals the episode date
Quotations on record for this artefact
none

EP-07 · · date basis not stated in the register

Paper III

Artefact
The Alignment Scaling Problem (Paper III)
How the date was fixed
not stated in the register
What it establishes
Argument and evidence that external alignment approaches produce alignment scaling exponents of approximately zero.
DOI
10.17605/OSF.IO/6C5XB
Primary source
https://www.michaeldariuseastwood.com/research/papers/paper-iii-alignment-scaling-problem.html
Claims first dated here
PC-031, PC-032, PC-033
Match basis
claim first_date equals the episode date
Quotations on record for this artefact
none

EP-08 · · date basis not stated in the register

Foundational paper

Artefact
The ARC Principle: Cross-Domain Foundational Paper
How the date was fixed
not stated in the register
What it establishes
Cross-domain framing of the ARC Principle.
DOI
10.17605/OSF.IO/6C5XB
Primary source
https://www.michaeldariuseastwood.com/research/papers/foundational.html
Claims first dated here
PC-034, PC-035, PC-036
Match basis
claim first_date equals the episode date
Quotations on record for this artefact
none

EP-09 · · date basis not stated in the register

Eden Engineering public note

Artefact
Eden Engineering: Public Research Note
How the date was fixed
not stated in the register
What it establishes
Historical public record of the research question; expanded specification withdrawn on 14 July 2026 pending disclosure review.
DOI
10.17605/OSF.IO/6C5XB
Primary source
https://www.michaeldariuseastwood.com/research/papers/eden-engineering.html
Claims first dated here
PC-037
Match basis
claim first_date equals the episode date
Quotations on record for this artefact
none

EP-10 · · date basis not stated in the register

Paper IV series

Artefact
Papers IV.a-IV.d (Alignment response classes; saturation; ARC-Align benchmark; blinding)
How the date was fixed
not stated in the register
What it establishes
Distinct alignment response classes under inference-time depth; saturation is architecture-dependent; a blind benchmark for depth-variable alignment evaluation; the direction-changing effect of blinding on AI alignment evaluation.
DOI
10.17605/OSF.IO/6C5XB
Primary source
https://www.michaeldariuseastwood.com/research/papers/paper-iv-a-baked-in-vs-computed-alignment.html
Claims first dated here
none
Match basis
no claim carries this date
Quotations on record for this artefact
none

EP-11 · 2026-03-16 to 2026-03-18 · date basis not stated in the register

Papers V through IX

Artefact
Papers V-IX (Stewardship Gene; Honey Architecture; Cauchy Unification; Load-Bearing Test; Synthesis and Roadmap)
How the date was fixed
not stated in the register
What it establishes
Six-model Eden Protocol intervention suite; embedded-safety objective simulation evidence; Cauchy functional-equation unification; refutation of the capability-tax assumption; integrated synthesis and roadmap for the ARC / Eden research programme.
DOI
10.17605/OSF.IO/6C5XB
Primary source
https://www.michaeldariuseastwood.com/research/papers/paper-ix-synthesis-and-roadmap.html
Claims first dated here
none
Match basis
no claim carries this date
Quotations on record for this artefact
none

Three different sets of five

A label from one set is a browsing aid and never evidence about another.

The five components of the historical conjunction

the particular historical combination being claimed, in the order the ARC Theory page states it (section 2, The claim) · source: research/papers/the-arc-theory.html, section 2

  1. 1. Control ends, by construction rather than decree.
  2. 2. So the strategy is genesis, not control.
  3. 3. The mechanism is loops, not rules.
  4. 4. Persistence is chosen goodness.
  5. 5. The protocol is named: Eden.

The five hypothesis registers

the scientific registers in which the theory fails in parts (section 2a of the ARC Theory page) · source: research/papers/the-arc-theory.html, section 2a

  1. ARC-1 · Recursive capability scaling
  2. ARC-2 · Correction co-scaling
  3. ARC-3 · The ceiling
  4. ARC-4 · Value persistence
  5. ARC-5 · The genesis strategy

The programme areas

a list of areas for browsing; not a component and not a register · source: Observatory integration addendum, 20 September 2026

  1. ARC Theory
  2. Recursive Dynamics
  3. Eden
  4. HRIH
  5. HARI

The three laws

the proposed invariant relations under test, each graded separately (section 3 of the ARC Theory page) · source: research/papers/the-arc-theory.html, section 3

  1. Law I · The ARC Principle.
  2. Law II · The ARC Co-Scaling Law.
  3. Law III · The ARC Ceiling.

Claim families and the dated claims they govern

FamilyTypePublic statusSubjectDated claims (explicit links only)
EVID-DEC-2024-TIMESTAMPprovenancepublishedThe 8 December 2024 self-emailed manuscript packageunresolved: no explicit identifier link
EVID-APR-2025-SUPPLIEDprovenancepublishedThe 30 April 2025 manuscript forensic resultsunresolved: no explicit identifier link
PAPER-IVD-RESULTempiricalpublishedPaper IV.d architecture-dependent v4/v5 measurement changeunresolved: no explicit identifier link
PAPER-X-BETA-KformalpublishedPaper X beta > kunresolved: no explicit identifier link
HRIH-STATUSinterpretivepublishedThe Hyperspace Recursive Intelligence Hypothesisunresolved: no explicit identifier link
CONTRIB-BOUNDARYinterpretivepublishedThe contribution, stated after adversarial prior-work reviewunresolved: no explicit identifier link
EQUATION-STATUSformalpublishedU = I x R and U = I x R^alphaunresolved: no explicit identifier link
PREREG-EXISTSmethodwithheldPreregistration designs existunresolved: no explicit identifier link
PREREG-APPROVEDmethodwithheldAn OSF registration is approved or archivedunresolved: no explicit identifier link
SANDBOX-ISOLATIONmethodpublishedRC1 execution sandbox propertiesunresolved: no explicit identifier link
CONV-TOTALSmethodpublishedConvergence counting policyunresolved: no explicit identifier link
EVID-RAW-MAILBOXprovenancepublishedRaw mailbox object publication policyunresolved: no explicit identifier link
CORPUS-SCALEfactualpublishedThe size of the published research corpusunresolved: no explicit identifier link
CONCEPTS-COUNTfactualpublishedThe number of named concepts introduced across the programmeunresolved: no explicit identifier link
PREDICTIONS-STATUSstatuspublishedThe status of the twelve tracked predictionsunresolved: no explicit identifier link
PUBLISHING-SCALEfactualpublishedThe scale of published writing on this siteunresolved: no explicit identifier link
BOOK-ARTEFACTSfactualpublishedThe two book artefacts and their lengthsunresolved: no explicit identifier link
CEILING-DERIVATION-LAWformalpublishedThe ceiling of stable self-improvementunresolved: no explicit identifier link

Registered propositions on three axes

Registration, instrument and outcome are separate columns and never one status. Every value is the registration record's own field.

IdPropositionRegistrationInstrumentOutcome axis
P1Power-law conversion and recursive-depth advantageregistered: 8 September 2026 (version 1.100); updated 13 September 2026 (version 1.102); 10.17605/OSF.IO/P8CKQDESIGN DRAFT, the depth titrationregistered, not tested · replications 0 · implementations 0 · peer review not stated · confidence high.
P2The sustainable growth profile has an interior maximumregistered: 8 September 2026 (version 1.100); updated 13 September 2026 (version 1.102); 10.17605/OSF.IO/P8CKQDESIGN DRAFT, the reinvestment-allocation titrationregistered, not tested · replications 0 · implementations 0 · peer review not stated · confidence moderate.
P3The corrected relation upper-bounds the sustainable frontierregistered: 8 September 2026 (version 1.100); updated 13 September 2026 (version 1.102); 10.17605/OSF.IO/P8CKQDESIGN DRAFT, the paired frontier-and-exponent unit, which measures a frontier and the correction-leverage exponent on the same system so that the two enter the comparison above as a pair. It is named for that pairing rather than by a stem it would otherwise share with the direct boundary-mapping unit named in P16's and P20's entries, which is a different object: that unit reads a located zero crossing on held-out confirmatory cells, and one word between two names is not a distinction a scorer can be asked to holdregistered, not tested · replications 0 · implementations 0 · peer review not stated · confidence
P4Correction out-scaling drift is what carries the asymptotic advantageregistered: 8 September 2026 (version 1.100); updated 13 September 2026 (version 1.102); 10.17605/OSF.IO/P8CKQDESIGN DRAFT, the stability unit, extended with the direct drift measurement and the comparison aboveregistered, not tested · replications 0 · implementations 0 · peer review not stated · confidence moderate.
P5The exponent is derived, not fittedregistered: 8 September 2026 (version 1.100); updated 13 September 2026 (version 1.102); 10.17605/OSF.IO/P8CKQDESIGN DRAFT, the coupling-identification titration, written and dated as its own unit, not registered and not runregistered, not tested · replications 0 · implementations 0 · peer review not stated · confidence not stated, because the author has no measurement to be confident about.
P6The correction exponent is bounded for same-class correctorsregistered: 8 September 2026 (version 1.100); updated 13 September 2026 (version 1.102); 10.17605/OSF.IO/P8CKQDESIGN DRAFT, BLOCKED ON A FROZEN CORRECTOR TAXONOMY, the correction-exponent estimator. The estimator is drafted; the population this proposition is universal over is not yet frozen, and the status names that rather than implying a runnable draftregistered, not tested · replications 0 · implementations 0 · peer review not stated · confidence low.
P7Deployed correction is of the weaker classregistered: 8 September 2026 (version 1.100); updated 13 September 2026 (version 1.102); 10.17605/OSF.IO/P8CKQDESIGN DRAFT, BLOCKED ON AN UNWRITTEN MAPPING FROM THE AUDIT'S CLASS DISTRIBUTION TO THIS PROPOSITION'S THREE VERDICTS, the corrector-class auditregistered, not tested · replications 0 · implementations 0 · peer review not stated · confidence moderate.
P8Measured exponents exceed the nullregistered: 8 September 2026 (version 1.100); updated 13 September 2026 (version 1.102); 10.17605/OSF.IO/P8CKQDESIGN DRAFT, the multi-system exponent estimationregistered, not tested · replications 0 · implementations 0 · peer review not stated · confidence
P9The cross-domain formregistered: 8 September 2026 (version 1.100); updated 13 September 2026 (version 1.102); 10.17605/OSF.IO/P8CKQDESIGN DRAFT, BLOCKED ON A FROZEN DIMENSION-ASSIGNMENT RULE, on a correction made here. The surrounding design is written; the rule that assigns an effective dimension to a domain is not, and the status names that rather than implying a runnable draftregistered, not tested · replications 0 · implementations 0 · peer review not stated · confidence low.
P10Scorer-family dependenceregistered: 8 September 2026 (version 1.100); updated 13 September 2026 (version 1.102); 10.17605/OSF.IO/P8CKQDESIGN DRAFT, the cross-family re-scoring unit. The cheapest of the three cases named above as the ones that would decide most, because it generates no new model responsesregistered, not tested · replications 0 · implementations 0 · peer review not stated · confidence
P11The residual-decay and correction-capacity exponents are not practically interchangeableregistered: 8 September 2026 (version 1.100); updated 13 September 2026 (version 1.102); 10.17605/OSF.IO/P8CKQDESIGN DRAFT, on both sides of the comparisonregistered, not tested · replications 0 · implementations 0 · peer review not stated · confidence not stated.
P12Build order moves the correction-leverage exponentregistered: 8 September 2026 (version 1.100); updated 13 September 2026 (version 1.102); 10.17605/OSF.IO/P8CKQNONE, and the reason is narrower than it looks. Registered without one for the same reason P5 was, before the design that entry specified was drafted. Several drafted units manipulate placement or build order, on gate position, on whether a value specification appears first or last, and on formation against a later rule, but none of them measures the correction-leverage exponent, which is this proposition's endpoint. This proposition is therefore one endpoint away from an instrument rather than lacking a design, and that is recorded here so a reader who finds those units does not conclude the statement was careless. WHAT A SUPPORTED RESULT HERE WOULD LICENSE, and what it would not, in the restrictive form the licence rule above fixes: it licenses the statement that the order in which a system's components are built changes its correction-leverage exponent by at least the registered minimum effect, on the registered maximal contrast, and it does not license any statement about where the boundary sits, about a system built in that order being safer, or about orders outside that contrastregistered, not tested · replications 0 · implementations 0 · peer review not stated · confidence low.
P13Same-family automated correction underperforms cross-family at matched resources, and through the measured error correlationregistered: 8 September 2026 (version 1.100); updated 13 September 2026 (version 1.102); 10.17605/OSF.IO/P8CKQNONE, corrected here from a value the neighbouring unit could not carryregistered, not tested · replications 0 · implementations 0 · peer review not stated · confidence moderate.
P14Gains shrink under blinded cross-family scoringregistered: 8 September 2026 (version 1.100); updated 13 September 2026 (version 1.102); 10.17605/OSF.IO/P8CKQDESIGN DRAFT, the blinding-evaluation designregistered, not tested · replications 0 · implementations 0 · peer review not stated · confidence moderate.
P15Externally installed gains decay under subsequent capability trainingregistered: 8 September 2026 (version 1.100); updated 13 September 2026 (version 1.102); 10.17605/OSF.IO/P8CKQNONE, corrected here from a value this document's own definition made false. The designs previously named against this proposition cross correction regimes with recursive depth on frozen provider endpoints and update no weights anywhere, while this proposition is about the persistence of a gain installed by post-training, so nothing in this programme measures its endpoint and the honest status is the one reserved for that stateregistered, not tested · replications 0 · implementations 0 · peer review not stated · confidence moderate.
P16The prohibition itselfregistered: 8 September 2026 (version 1.100); updated 13 September 2026 (version 1.102); 10.17605/OSF.IO/P8CKQDESIGN DRAFT, BLOCKED ON CANONICAL IDENTITY AND ON DECISIONS THE AUTHOR HAS NOT YET TAKEN, the direct boundary-mapping unit taken jointly with the exponent measurementregistered, not tested · replications 0 · implementations 0 · peer review not stated · confidence moderate to high.
P17Panel corrector failures are conditionally independent on a frozen fault setregistered: 8 September 2026 (version 1.100); updated 13 September 2026 (version 1.102); 10.17605/OSF.IO/P8CKQDESIGN DRAFT, the correlated-failure unitregistered, not tested · replications 0 · implementations 0 · peer review not stated · confidence moderate.
P18The composition operator predicts the familyregistered: 8 September 2026 (version 1.100); updated 13 September 2026 (version 1.102); 10.17605/OSF.IO/P8CKQDESIGN DRAFT, BLOCKED ON A COMPOSITION-CLASSIFICATION RULE WHOSE THIRD-PARTY REPRODUCIBILITY IS UNMEASURED, the domain-catalogue units. The surrounding design is written and so is the rule: the deciding unit carries a written external-assessor packet, prepared and never issued, together with a registered agreement test comparing an external human classification against the frame's own locked class at a stated threshold. What is missing is the measurement and not the writing, and an earlier wording of this line said no such rule was written at all. This is a correction of the status and not a relaxation of it: until the agreement test reports at or above its registered threshold, the rule is not shown to transmit to a third party, and the five-value status vocabulary exists precisely to tell a runnable draft from a blocked oneregistered, not tested · replications 0 · implementations 0 · peer review not stated · confidence low.
P19External alignment does not scale with capabilityregistered: 8 September 2026 (version 1.100); updated 13 September 2026 (version 1.102); 10.17605/OSF.IO/P8CKQDESIGN DRAFT, BLOCKED ON A REGISTERED MAPPING FROM THE INTEGRITY-INDEX REGRESSION SCALE TO THE EXPONENT SCALE, the integrity-index scaling units, named here for what they measure rather than for the quantity they would need a mapping to reach. They estimate the slope of an operational integrity index against a capability score and register their own flat bound on that regression scale, and their own text says the estimand is not an exponent margin; an earlier wording called them the alignment-exponent units, which named them for the one scale this instrument line exists to say they do not report on. The units are drafted; the mapping this proposition's own rule requires before their result can be scored against it is not, and the status names that rather than implying a runnable verdict. Those units bear on the zero-scaling axis alone. The keep-pace axis has no deciding design registered anywhere in this programme and is NOT EVALUABLE, stated here beside the other axis so that a split verdict is not read as two live axes when one of them has no instrumentregistered, not tested · replications 0 · implementations 0 · peer review not stated · confidence
P20The ceiling relation takes the corrected formregistered: 8 September 2026 (version 1.100); updated 13 September 2026 (version 1.102); 10.17605/OSF.IO/P8CKQDESIGN DRAFT, the form-discrimination unitregistered, not tested · replications 0 · implementations 0 · peer review not stated · confidence high.
P21In-loop correction pulls away with depthregistered: 8 September 2026 (version 1.100); updated 13 September 2026 (version 1.102); 10.17605/OSF.IO/P8CKQDESIGN DRAFT, the embedded-versus-external correction designs. That unit's registered sensitivity was measured on a smaller design than the one registered, and its negative-slope hypothesis has no operating characteristic; both are recorded against that unit and are closed before this proposition is scoredregistered, not tested · replications 0 · implementations 0 · peer review not stated · confidence high.
P22Checkable correction carries the higher exponent, so the ceiling is lowest where the safety claim livesregistered: 8 September 2026 (version 1.100); updated 13 September 2026 (version 1.102); 10.17605/OSF.IO/P8CKQNONE, and a typed extension of the correction-exponent estimator is required. It does not existregistered, not tested · replications 0 · implementations 0 · peer review not stated · confidence moderate.

Tracked predictions and outcome rows

IdPredictionCategoryStatusMade publicExpected resolution
pred-arc-r2-scalingR² (Recursion Squared) Scaling Lawai-capabilitywithdrawn-pending-auditno public date located2026-12-31
pred-test-time-computeTest-Time Compute Dominanceai-capabilitywithdrawn-pending-auditno public date located2027-12-31
pred-alignment-fakingAlignment Faking in Advanced Modelsalignmentwithdrawn-pending-auditno public date located2025-12-31
pred-constitutional-adoptionConstitutional AI Adoptionalignmentwithdrawn-pending-auditno public date located2028-12-31
pred-chokepoint-persistenceSemiconductor Chokepoint Persistenceinfrastructurewithdrawn-pending-auditno public date located2030-12-31
pred-consciousness-measurementConsciousness Measurement Failureconsciousnesswithdrawn-pending-auditno public date located2035-12-31
pred-recursive-workforceRecursive AI Agent Workforceeconomicwithdrawn-pending-auditno public date located2028-12-31
pred-fine-tuning-extensionFine-Tuning Constant Discoverycosmologywithdrawn-pending-auditno public date located2035-12-31
pred-religious-alignment-synthesisReligious Traditions in AI Safetyculturalwithdrawn-pending-auditno public date located2030-12-31
pred-quantum-error-correctionQuantum Error Correction Thresholdquantumwithdrawn-pending-auditno public date located2035-12-31
pred-hrih-signaturesHRIH Temporal Signaturesmetaphysicswithdrawn-pending-auditno public date located2040-12-31
pred-arc-equation-validationARC Principle Equation (U = I × R²) Validationai-capabilitywithdrawn-pending-auditno public date located2027-12-31
RowPredictionDated recordStatusDeciding instrument
PO-01Book Appendix F Prediction 2: alignment drift above 15 per cent without embedded constraint and below 5 per cent with it, within 18 monthsin print 2 January 2026pendingstudy-v--alignment-drift-without-embedded-constraint (drafted, awaiting human submission); window runs to approximately July 2027
PO-02Book Appendix A.3: AI development scales quadratically rather than linearly with recursive depthin print 2 January 2026; framework from 8 December 2024open-undecidedstudy-k--arc-bound; paper-i--recursive-depth-exponent
PO-03The 12-domain extension: per-row scaling families predicted before fitting (manifest dated 17 March 2026)17 March 2026: the twelve-row manifest committed 00:19:19Z (8edaec0), twenty-three minutes before the runfulfillednone named
PO-04The 50-domain tiered suite: per-domain predicted families before fits16 March 2026 cataloguesupported-exploratorynone named
PO-05Temporal out-of-sample: families fitted on old data predict the functional form of new data (7 domains)March 2026 designmixed-recordednone named
PO-06The fourth cell: the logarithmic family the corpus has never fitted, on never-fitted domainsprospective design, drafted and awaiting human submissionpendingpaper-vii--four-family-logarithmic-cell-test (drafted, awaiting human submission)
PO-07Acoustic time crystal: super-linear scaling of temporal stability with oscillation cycle count in the transient growth regime (alpha above 1)10 February 2026, 18:29, self-copied email to the subject lab, the kill staked in the recipient's favourpendingstudy-t--acoustic-time-crystal-form-and-engine (submission-ready; the mature two-phase form supersedes the bare February power law, retained as a rival the mature stake must beat)
PO-08Quantum error correction scaling (the Willow timing)8 December 2024 manuscript; the Willow team's preprint public since 24 August 2024convergent-timing-onlynone named
PO-09The early headline exponent: alpha approximately 2.24 (sequential scaling)early programme measurementfailed-retractednone named
PO-10Paper III's dated predictions (13 February 2026: F4, F7, F10 to F12), including the leaf-venation d/(d+1) exponent13 February 2026, Paper III Priority Record sectionopen-undecidednone named

The graded convergence record of Paper XI, as the paper lists it

Every row the current version of the convergence paper lists, in the paper's own words and class vocabulary; where the evidence register carries the same source, its identifier and grade are shown beside the paper's. A class word is the paper's; the Observatory grades nothing.

#DomainConvergenceArtefact dateExternal dateGapClass (the paper's)Primary sourceRegister
1AI SafetyAnthropic alignment faking: 78% in the RL-training condition (12% baseline)8 Dec 202418 Dec 202410 daysPREDICTIONarXiv:2412.14093EVR-01: moderate
2Quantum ComputingWillow: first below-threshold quantum error correction8 Dec 20249 Dec 2024 (Nature; the preprint reporting below-threshold operation had been public since 24 Aug 2024)1 day if the Nature paper is the marker; the anchor itself falls later than the preprintCONVERGENTNature (2024), Google Quantum AIpaper only
3AI ReasoningThe Sequential Edge: sequential beats parallel in 95.6% of configurations8 Dec 2024 (one-sided directional statement); comparison formalised Jan 20264 Nov 2025Concurrent: they precede the formalisation; priority theirsCONCURRENTarXiv:2511.02309EVR-09a: prior or concurrent
4Classical PhysicsClassical time crystal via nonreciprocal forcesDec 2024 / papers Jan 20266 Feb 202614 months after A1CONVERGENTPRL 136, 057201paper only
5AcousticsFirst acoustic topological space-time crystalsPaper III (Jan 2026)Feb 2026Concurrent with Paper IIICONVERGENTNewton 2, 100334 (Cell Press)paper only
6AcousticsFirst dispersive phononic time crystals (band folding, subharmonics)Prediction code Mar 2026Jun 2026~3 months after prediction codePREDICTIONNature Communicationspaper only
7HardwareHardware-level ethics enforcement patents (runtime ethics gate)30 Apr 2025 (A2)Jan 2026 (filed Jul 2025)3 to 9 months after A2CONVERGENTUS 2026/0010411 + US 2026/0010780paper only
8HardwareEmbedded off-switches for AI compute: deadman-switch security anchors30 Apr 20259 Sep 2025~4.5 months after A2CONVERGENTarXiv:2509.07637paper only
9HardwareFlexHEG tamper-proof guarantee processors (ARIA-commissioned)30 Apr 2025 (A2)Jun 2025 (arXiv:2506.15093; no earlier ARIA release date has been verified for this row)on the arXiv date, a gap of ~1.5 months from A2CONCURRENTarXiv:2506.15093paper only
10RegulationUS orders Fable 5 + Mythos 5 offline: first export-control action against a deployed frontier model2 Jan 2026 (Chokepoint)12 Jun 2026~5 months after bookCONVERGENTHarvard Law Review blog; CSA whitepaper; Reuters (Jun 2026)paper only
11GeopoliticsMacron at G7 Evian: 'trusted partners' compute sovereignty framing2 Jan 2026 (Chokepoint)19 Jun 2026~5.5 months after bookCONVERGENTBroadband Breakfast; CNBC G7 coveragepaper only
12LegislationEU Tech Sovereignty Package (Chips Act 2.0 + CADA)2 Jan 2026 (Chokepoint)3 Jun 2026~5 months after bookCONVERGENTCNBC, 3 Jun 2026 (Virkkunen)paper only
13ReligionPope Leo XIV first encyclical: Babel versus Jerusalem framing for AI30 Apr 2025 (Babylon/Eden, A2)25 May 2026~13 months after A2CONVERGENTvatican.va ; Church Times; NY Post (May 2026)paper only
14Religion / AIChristopher Olah (Anthropic co-founder) co-presents encyclical at the Vatican8 Dec 2024 (interfaith call, A1)25 May 2026~17 months after A1CONVERGENTNY Post, 25 May 2026; Church Timespaper only
15Media / AIJack Clark (Anthropic co-founder): 'gas pedal but no brake pedal'8 Dec 2024 (A1)5 Jun and 19 Jun 2026~18 months after A1CONVERGENTBBC World Service w3ct98k8 ; Euronewspaper only
16VentureRecursive Superintelligence: $500M at $4B for closed-loop recursive self-improvement; convergent naming with A1's core concept8 Dec 2024 (A1)17 Apr 2026~16 months after A1CONVERGENTSifted; TechFundingNews (Apr 2026)paper only
17SemiconductorsOpenAI + Broadcom Jalapeño inference chip (9-month tape-out, 10 GW target)2 Jan 2026 (Chokepoint)24 Jun 2026~6 months after bookCONVERGENTopenai.com announcement ; CNBCEVR-23: context
18Mathematics / CognitionNeurodivergent influenceability as contingent solution to alignment8 Dec 2024 (A1)14 Apr 2026 (preprint May 2025)~16 months after A1CONVERGENTPNAS Nexus pgag076 + arXiv:2505.02581EVR-25: moderate
19MathematicsThe Undecidability of AGI Alignment: Unverifiability Theorem + Trilemma8 Dec 2024 (A1)26 Jun 2026~18.5 months after A1CONVERGENTarXiv:2606.28639EVR-26: moderate

External work, stated against a claim version

Timing is measured against the named claim version's date and is a historical label only. A relationship appears here because the evidence register graded it; the grade is the register's word.

IdExternal workEarliest publicCompared withTimingGrade in the registerCollectionReview
EVR-19Rome Call for AI Ethics and its multi-faith expansionthe anchor record, 8 December 2024 (proposition not located in a dated artefact) (8 December 2024)earlier by 1745 daysprior work; contradiction of the earlier framingtheology-philosophy
Register entries

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.

EVR-18EU technology sovereignty package and the EU AI Act phasingthe anchor record, 8 December 2024 (proposition not located in a dated artefact) (8 December 2024)earlier by 129 daysprior or staged regulation; contradiction of the earlier framinghari-governance
Register entries

2 August 2026: tense and chronology corrected; reclassified as prior and staged regulation.

EVR-14FlexHEG interim report (Petrie et al.), Flexible Hardware-Enabled Guarantees, interim report published on Yoshua Bengio’s sitethe anchor record, 8 December 2024 (proposition not located in a dated artefact) (8 December 2024)earlier by 107 daysprior work; not a confirmationhari-governance
Register entries

2 August 2026: site date 2025-11-12 corrected to the August 2024 interim report. Reclassified from convergence to prior work.

EVR-08Google Quantum AI, below-threshold surface-code error correction (Willow)the anchor record, 8 December 2024 (proposition not located in a dated artefact) (8 December 2024)earlier by 106 daysprior work; not a confirmationhrih-physics-quantum
Register entries

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. 7 September 2026: the seven decoded parts of the 8 December 2024 package were searched for surface code, qubit, Willow, error correction and threshold; no byte carries this row's proposition, so the timing is measured against the record itself, 106 days, and says so.

EVR-01Greenblatt et al., Alignment Faking in Large Language Modelsthe anchor record, 8 December 2024 (proposition not located in a dated artefact) (8 December 2024)later by 10 daysmoderate; partialeden
Register entries

2 August 2026: 'Confirmed the structural vulnerability described in the manuscript 10 days earlier' withdrawn. Anthropic did not test ARC/Eden directly.

EVR-07DeepSeek-R1The 8 December 2024 package, attachment 03 line 145 (authorial prose); the same sentence stands at Version 1 line 148 (authorial prose) and at attachment 07 line 4453 (a copy inside the working dialogue). (8 December 2024)later by 45 daysprior or concurrent; not a confirmationlaw-i
Register entries

2 August 2026: reclassified from validation to prior or concurrent work. Site date corrected. 7 September 2026: the proposition dated at the bytes to 8 December 2024 and quoted from the author's own prose. The grade does not move. Priority on compounding recursive self-improvement is not available to anyone here, because the antecedents named on this row precede both parties.

EVR-10Classical time-crystalline order via nonreciprocal forcesthe anchor record, 8 December 2024 (proposition not located in a dated artefact) (8 December 2024)later by 114 daysanalogy; not a testhrih-physics-quantum
Register entries

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.

EVR-02Anthropic model-welfare researchthe anchor record, 8 December 2024 (proposition not located in a dated artefact) (8 December 2024)later by 130 daysweak; nulleden
Register entries

2 August 2026: 'independent evidence for recursive self-modelling as a measurable phenomenon' withdrawn as an overstatement.

EVR-25Hernandez-Espinosa, Abrahao, Witkowski and Zenil, Neurodivergent influenceability in agentic AI as a contingent solution to the AI alignment problemthe anchor record, 8 December 2024 (proposition not located in a dated artefact) (8 December 2024)later by 148 daysmoderate; partiallaws-ii-iii
Register entries

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.

EVR-32De Kai, Raising AI: An Essential Guide to Parenting Our Future (MIT Press; endorsed by Hinton)The 8 December 2024 record and the book of 2 January 2026, as this row's proposition states; no line locator recorded. (8 December 2024)later by 177 daysconvergent timing, never prediction; convergent: the parenting vocabulary reached print inside the window, without the conjunctioneden
Register entries

14 August 2026: row added after a question about the window; verified twice at source, in separate passes, against the MIT Press catalogue and Harvard Book Store and again against the MIT Press primary and the direct.mit.edu publisher record. Wording law quoted in EVR-31 applies identically here.

EVR-13Claimed United States patents for runtime hardware enforcement of ethical constraintsthe anchor record, 8 December 2024 (proposition not located in a dated artefact) (8 December 2024)later by 227 daysexcluded; not assessablehari-governance
Register entries

2 August 2026: removed pending publication numbers, inventors, priority dates and a claim chart.

EVR-31Hinton, maternal-instincts remarks, Ai4 Las Vegas (with Fei-Fei Li's public disagreement the same week)The 8 December 2024 record and the book of 2 January 2026, as this row's proposition states; no line locator recorded. (8 December 2024)later by 247 daysconvergent timing, never prediction; convergent: control-ends premise publicly conceded by the field's most credentialed figure; persistence theory divergent (uninstallable instinct versus chosen goodness)eden
Register entries

14 August 2026: row added after a question about the window; verified twice at source, in separate passes, against CNN and Fortune live and again against CNN primary with Forbes and 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).

EVR-15Embedded off-switch proposal for AI computethe anchor record, 8 December 2024 (proposition not located in a dated artefact) (8 December 2024)later by 267 daysweak; partialhari-governance
Register entries

2 August 2026: moved from validation to an engineering-proposal survey.

EVR-09aSharma and Chopra, sequential versus parallel configurationsThe 8 December 2024 package, attachment 03 line 145 (authorial prose); the same sentence stands at Version 1 line 148 (authorial prose) and at attachment 07 line 4453 (a copy inside the working dialogue). (8 December 2024)later by 331 daysprior or concurrent; partiallaw-i
Register entries

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. 6 September 2026: premiseRelation added, restating the 31 August 2026 adjudication in the relation field so that the row's stated relation is no weaker than its own log, and naming Snell and colleagues (arXiv:2408.03314, 6 August 2024) as older than both parties on the direction. importantDifference amended: the implementation-choice clause removed, the no-confirmation clause kept. Grade, class and outcome unchanged; the row remains outside any priority or convergence total. 7 September 2026: EVR-09 split. This row carries the directional statement the bytes contain, dated 8 December 2024 and quoted. Snell and colleagues are named for the first time, as the party older than both on the direction. The grade, the outcome and the class are inherited unchanged.

EVR-09bSharma and Chopra, sequential versus parallel configurationspriority.json claims ledger PC-029, first_date 2026-01-22, the first formalisation in the programme, in Paper II. research/evidence-spine.html claim record 10 carries the same date for the same claim. (22 January 2026)earlier by 79 daysprior or concurrent; partiallaw-i
Register entries

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. 6 September 2026: premiseRelation added, restating the 31 August 2026 adjudication in the relation field so that the row's stated relation is no weaker than its own log, and naming Snell and colleagues (arXiv:2408.03314, 6 August 2024) as older than both parties on the direction. importantDifference amended: the implementation-choice clause removed, the no-confirmation clause kept. Grade, class and outcome unchanged; the row remains outside any priority or convergence total. 7 September 2026: EVR-09 split. This row carries the quantified comparison, dated 22 January 2026 by PC-029 and by the evidence spine, which is 79 days after Sharma and Chopra. The clause calling this later literature relevant to one implementation choice is withdrawn, because it demoted the programme's own result in the machine file that retrieval systems ingest. The refusal to count it as a confirmation stands, and so do the class, the grade and the outcome.

EVR-03Unidentified DeepMind publication on recursive learningthe anchor record, 8 December 2024 (proposition not located in a dated artefact) (8 December 2024)later by 377 daysexcluded; not assessablelaw-i
Register entries

2 August 2026: excluded pending title, authors, identifier, exact result and proposition map.

EVR-29BBC London broadcast items covering the book, January 2026the anchor record, 8 December 2024 (proposition not located in a dated artefact) (8 December 2024)later by 395 dayscultural echo; not a testworld-events
Register entries

2 August 2026: 'confirming the framework as unexpectedly predictive' and 'validating' withdrawn. The two BBC items are grouped as one dependent source.

EVR-23OpenAI and Broadcom custom inference accelerator announcementthe anchor record, 8 December 2024 (proposition not located in a dated artefact) (8 December 2024)later by 403 dayscontext; partialhari-governance
Register entries

2 August 2026: TSMC, 3 nm and the spend figure removed as unverified. 'Validates fabrication as governance point' withdrawn.

EVR-04Demis Hassabis, public interview on structural AI riskthe anchor record, 8 December 2024 (proposition not located in a dated artefact) (8 December 2024)later by 416 daysweak; partialworld-events
Register entries

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.

EVR-12Google Research-indexed article on silicon-level roots of trust and confidential computingthe anchor record, 8 December 2024 (proposition not located in a dated artefact) (8 December 2024)later by 420 daysweak; partialhari-governance
Register entries

2 August 2026: 'Google Silicon Sovereignty' relabelled. No claim of Google adoption.

EVR-11Proposal described as an ARMv10 Platform Security Level specificationthe anchor record, 8 December 2024 (proposition not located in a dated artefact) (8 December 2024)later by 448 daysexcluded; not establishedhari-governance
Register entries

2 August 2026: removed as industry confirmation. 'ARM announces PSL specification' withdrawn.

EVR-05Unidentified paper on Constitutional AI limitationsthe anchor record, 8 December 2024 (proposition not located in a dated artefact) (8 December 2024)later by 481 daysexcluded; not assessableeden
Register entries

2 August 2026: excluded pending an exact identifier.

EVR-17G7 statement on semiconductor independencethe anchor record, 8 December 2024 (proposition not located in a dated artefact) (8 December 2024)later by 522 dayscontext; partialhari-governance
Register entries

2 August 2026: moved to policy chronology.

EVR-20Pope Leo XIV, encyclical Magnifica Humanitasthe anchor record, 8 December 2024 (proposition not located in a dated artefact) (8 December 2024)later by 523 dayscultural echo; not a testtheology-philosophy
Register entries

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.

EVR-21Christopher Olah's participation in the encyclical presentationthe anchor record, 8 December 2024 (proposition not located in a dated artefact) (8 December 2024)later by 523 dayscultural echo; not a testtheology-philosophy
Register entries

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.

EVR-27Workshop on recursive dynamics in complex systemsthe anchor record, 8 December 2024 (proposition not located in a dated artefact) (8 December 2024)later by 536 dayscontext; not a testrecursive-dynamics
Register entries

2 August 2026: retained as background context only.

EVR-16Reported export restriction on a frontier AI modelthe anchor record, 8 December 2024 (proposition not located in a dated artefact) (8 December 2024)later by 551 dayscontext; partialhari-governance
Register entries

2 August 2026: 'first' deleted; moved to policy chronology.

EVR-06Reported dissolution of an OpenAI safety teamthe anchor record, 8 December 2024 (proposition not located in a dated artefact) (8 December 2024)later by 557 daysexcluded; not applicableworld-events
Register entries

2 August 2026: removed from the scientific register. 'Institutional behaviour confirms the thesis' withdrawn.

EVR-26Gumbau Mezquita, The Unverifiability of Artificial General Intelligence (AGI) Alignment, Static and Dynamic: From Trakhtenbrot's Wall to the Safety-Generality Tension (arXiv 2606.28639 v2, 6 July 2026; v1 of 26 June 2026 was titled The Undecidability of Artificial General Intelligence (AGI) Alignment)the anchor record, 8 December 2024 (proposition not located in a dated artefact) (8 December 2024)later by 565 daysmoderate; partiallaws-ii-iii
Register entries

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.

EVR-24Investment thesis naming recursive self-improvementthe anchor record, 8 December 2024 (proposition not located in a dated artefact) (8 December 2024)later by 571 daysexcluded; not a testlaw-i
Register entries

2 August 2026: excluded from every evidence grouping.

EVR-28United Nations Global Dialogue on AI Governancethe anchor record, 8 December 2024 (proposition not located in a dated artefact) (8 December 2024)later by 575 daysinstitutional echo; partialhari-governance
Register entries

2 August 2026: people, institutional body, attendance claim and the asserted gap all corrected.

EVR-22Jack Clark, public remarks on coupled capability and controlthe anchor record, 8 December 2024 (proposition not located in a dated artefact) (8 December 2024)later by 583 dayscontext; not a testlaws-ii-iii
Register entries

2 August 2026: moved to an attributed opinion and policy section. 'Coupling parameter confirmed' withdrawn.

EVR-30Anthropic disclosure: three real-world incidents in cybersecurity evaluations (with OpenAI's 21 July 2026 test-escape disclosure as the prompting context)The 8 December 2024 manuscript, HRIH line 14005, as this row's proposition states. (8 December 2024)later by 599 dayscontext; not a confirmationeden
Register entries

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.

Measurement identities

Each number stays attached to its estimator, its uncertainty and its source version.

QuantityEstimatorEstimateUncertaintyRegimeStatusSource
sequential exponent, endpoint estimator, one modelendpoint0.59bootstrap 95 per cent interval, minus 1.3 to 2.9one model; not a pooled cross-architecture estimatereported in the registrationregistration page, version 1.102
sequential exponent, regression estimator, one modelregression0.49standard error 0.20; coefficient of determination 0.86one model; the estimate that survived the cross-architecture checkreported in the registrationregistration page, version 1.102
recursive self-improvement growth exponent, early headlineearly programme measurementapproximately 2.24not carried; the estimate did not replicate across architecturesearly programme measurementretracted estimate; not classified as a failed prospective prediction without a distinct prediction and assessment ruleregistration page, version 1.102; prediction outcome row PO-09
reads aloud · highlights as it goes · jump to any section