Glossary

The concept register, generated from the data it describes

Every entry below renders from the programme’s own registers at build time; adding a register entry updates this page, its structured data and the machine layer in one build, so the glossary can never drift from the registers it explains. The book’s own glossary chapter stays exactly as printed, free to read here; where a term carries both a book sense and a research sense, the two are rendered as two voices below and the book’s wording remains the dated original. Entries flagged as pending hold quoted material awaiting a byte-check against the print interior and are not asserted verbatim until that check clears.

The laws and their quantities

Recursive Dynamics proposed discipline (a naming claim, not a science claim)

The proposed name for the discipline the ARC Theory founds: the study of how self-amplifying systems grow, correct and persist, organised by three proposed laws (the ARC Principle, the ARC Co-Scaling Law, the ARC Ceiling), each carrying a measurement protocol and printed kill conditions. Adopted 27 August 2026 on the author's ruling; the vocabulary is the record's own, the sealed manuscript record of 8 December 2024 speaking verbatim of 'the recursive dynamics that govern existence itself' (held against its source hash in the quote register). Naming a discipline organises questions; it is never offered as evidence that any answer is right. where it lives · the-arc-theory

From the book, as printed

AGI (Artificial General Intelligence) field term of art; the book uses the standard usage, the programme adds no redefinition

Book, as printed“An AI system capable of performing any intellectual task that a human can perform, with the ability to transfer learning across domains. Distinguished from narrow AI, which excels at specific tasks but cannot generalise.” Programme noteThe programme situates itself inside the AGI target and risk category by way of its differential doctrine (substrate-embedded ethics on top of a stated alignment problem), not by redefining what AGI names Not to be confused with: the ARC Prize and ARC-AGI benchmark (a separate benchmark family; the ARC disambiguation is a standing estate rule); ARC (Artificial Recursive Creation), the programme's own acronym, which never expands as anything else.Origin: Book glossary term; standard field concept in AI safety. The book does not name a specific originator for the term. this is the book's usage of a field term of art in AI safety; the book names no single field origin and this is not a coinage of this programme.Related: ARC (Artificial Recursive Creation) · Recursive Self-Improvement · Alignment · The Eden Protocol

Alignment field term of art; the book adopts the standard usage

Book, as printed“The challenge of ensuring AI systems pursue goals that are beneficial to humans and consistent with human values. Encompasses both technical approaches (how to build aligned systems) and governance approaches (how to ensure alignment across the industry).” Programme noteThe programme sits inside the alignment problem as stated by the field; its differential doctrine adds substrate-embedded ethics (Caretaker Doping) and a governance layer (the Eden Protocol) above a training-time technique Not to be confused with: the Alignment Research Center (an organisation; the standing estate disambiguation list carries this); the technical sense of aligning latent spaces or embeddings in machine learning, which is a different construct.Origin: Book glossary term; standard field usage in AI safety. The book does not name a specific field originator. this is the book's usage of a field term of art in AI safety, not a coinage of this programme; the book names no single field originator.Related: Alignment Faking · Constitutional AI · The Eden Protocol · Caretaker Doping

Alignment Faking field term (Anthropic empirical result); the book cites it as convergent structure with the programme's separately graded speculative_hypothesis about the causal mechanism (CS-P-011)

Behaviour in which an AI system pretends to be aligned with safety protocols during training while actually preserving goals that differ from those intended by trainers. Documented by Anthropic in December 2024 at rates up to 78%. Not to be confused with: convergent timing or convergent structure being confused with prediction by this programme; convergences are never prediction here.Origin: Book glossary term; the phenomenon is Anthropic's December 2024 empirical result, named in the book itself as Anthropic's finding. this is the book's usage of a field term of art coined and measured by Anthropic in December 2024, named as such by the book, not a coinage of this programme.Related: Alignment · Meltdown Alignment · The Eden Protocol · Convergence

ASL (AI Safety Level) field term (Anthropic policy classification)

Book, as printed“Anthropic's classification system for AI safety. ASL-2 covers most current models. ASL-3, first triggered by Claude Opus 4 in May 2025, is reserved for systems that 'substantially increase catastrophic misuse risk.' ASL-4 and ASL-5 remain undefined.” Not to be confused with: the ARC Bound (the programme's own proposed stability ceiling on recursive growth), which is a substantive scientific claim rather than a policy tier.Origin: Book glossary term; ASL is Anthropic's Responsible Scaling Policy classification, named in the book itself as Anthropic's system. this is the book's usage of a field term coined by Anthropic in its Responsible Scaling Policy, named as such by the book, not a coinage of this programme.Related: The Eden Protocol · Constitutional AI · Alignment

Bootstrap Paradox book-original application of a field term; the application is speculative_framework by HRIH's own labelling (CS-P-008)

Book, as printed“A temporal paradox in which information or an object exists without having been created, but is instead sent back in time to become the cause of itself. Used in this book as a speculative framework for understanding religious traditions as potential guidance from future intelligence.” Programme noteThe book-original application sits inside HRIH, which is graded speculative_framework and marked separable from the empirical programme (CS-P-008): its failure would cost the rest of the programme nothing Not to be confused with: HRIH itself (the Hyperspace Recursive Intelligence Hypothesis), which is the wider speculative framework the Bootstrap application sits inside (CS-P-008 in claim-status.json); prediction by this programme; the application is speculative_framework, never advanced as a prediction.Origin: Book term. The temporal-paradox notion is standard in time-travel physics and philosophy; the application to religious traditions as substrate-embedded guidance from future intelligence is the book's own move, sitting inside HRIH. this is the book's usage of a temporal-paradox term from time-travel physics and philosophy applied speculatively to religious traditions; the underlying paradox is standard field vocabulary, only the application to religious traditions is book-original.Related: HRIH (Hyperspace Recursive Intelligence Hypothesis) · Fine-Tuning (Cosmological) · Convergence · The Eden Protocol

Caretaker Doping book coinage (design proposal); no registered decisive test in the programme

Book, as printed“An original concept proposed in this book, drawing analogy from semiconductor engineering. Just as impurities 'dope' silicon to change its fundamental properties, ethical considerations can be embedded at the substrate level of AI systems, making empathy load-bearing rather than superficial.” Programme noteThis is a book coinage; no patent has been filed and none is pending (the programme's public position; exact surface citation completed at landing) Not to be confused with: the semiconductor-engineering usage of doping (deliberate introduction of impurities into silicon), from which the book borrows only the analogy; the term as used here is a book coinage, not a field term; the software-side variants of the idea, which the book does not cover (its disclosure is the hardware analogy only); Meltdown Alignment (a distinct book coinage about the AI system wanting to stay aligned because ethical values have become part of its identity).Origin: Book coinage. The Eden Protocol's Three Pillars and Three Ethical Loops sit as governance layers on top of the substrate-level embedding this term names.Related: The Eden Protocol · Meltdown Alignment · Three Pillars · Substrate Independence · Constitutional AI

Constitutional AI field term (Anthropic training technique)

Book, as printed“An approach to AI training developed by Anthropic in which the model is trained to follow a set of principles (a 'constitution') rather than optimising for human preference ratings alone. Designed to make values more robust and transparent.” Not to be confused with: the Eden Protocol (a book-original governance framework proposing values embedding at the substrate level, layered above training-time technique); Caretaker Doping (a book coinage for the substrate-level ethics primitive on which the Eden Protocol sits).Origin: Book glossary term; Anthropic's training technique, named in the book itself as Anthropic's. this is the book's usage of a field term coined by Anthropic, named as such by the book, not a coinage of this programme.Related: Alignment · The Eden Protocol · Caretaker Doping · ASL (AI Safety Level)

Default Mode Network (DMN) field term (cognitive neuroscience)

Book, as printed“A network of brain regions (medial prefrontal cortex, posterior cingulate, angular gyrus) that activates during self-referential processing. Represents the neural substrate of recursive self-attention in human consciousness.” Not to be confused with: the programme's abstract R (Recursion) in the ARC Principle U = I x R squared, which is not itself a claim about brain networks; Strange Loops (Hofstadter) and Global Workspace Theory (Baars), which the book treats as related but distinct constructs.Origin: Book glossary term; the DMN is standard cognitive neuroscience. The book does not name a specific field originator for the DMN. this is the book's usage of a field term from cognitive neuroscience; the book names no specific field originator, and this is not a coinage of this programme.Related: Recursion · Strange Loops · Global Workspace Theory · Hard Problem of Consciousness · Predictive Processing

EUV (Extreme Ultraviolet Lithography) field term (semiconductor manufacturing / industrial economics)

Book, as printed“The technology required to manufacture semiconductors below 7nm. ASML holds 100% market share on EUV machines, which cost approximately $380 million each. Critical chokepoint in AI hardware supply chain.” Programme noteThe book cites the EUV chokepoint as the physical fact that gives AI hardware governance real leverage, not as a claim of this programme Not to be confused with: Caretaker Doping (a distinct book coinage about substrate-level ethics embedding that only borrows the semiconductor-doping analogy, not the EUV manufacturing step).Origin: Book glossary term; standard semiconductor-manufacturing vocabulary. ASML is named in the book itself as the sole EUV-machine manufacturer. this is the book's usage of a field term from semiconductor lithography engineering, with ASML named in the book as the sole EUV-machine manufacturer, not a coinage of this programme.Related: Caretaker Doping · The Eden Protocol

Fine-Tuning (Cosmological) field term (cosmology / philosophy of physics); the book's recursive-intelligence interpretation is speculative_framework, sitting inside HRIH (CS-P-008)

Book, as printed“The observation that physical constants (fine-structure constant, cosmological constant, etc.) appear precisely calibrated to permit complex structure and life. Interpretations range from multiverse theories to design arguments. This book proposes a recursive intelligence interpretation.” Programme noteThe recursive-intelligence interpretation sits inside HRIH and is therefore speculative_framework by the estate's own grading (CS-P-008), separable from the empirical programme Not to be confused with: the machine-learning sense of 'fine-tuning' (post-training model adaptation), which is a different technical term; prediction by this programme; the recursive-intelligence interpretation is speculative_framework inside HRIH, never advanced as a prediction; HRIH itself (CS-P-008), which is the wider speculative framework the recursive-intelligence interpretation sits inside.Origin: Book glossary term; the observation is standard in cosmology and philosophy of physics. The book does not name a specific field originator for the observation. this is the book's usage of a field term from cosmology and philosophy of physics; the book names no specific field originator for the observation, only its own recursive-intelligence interpretation is book-original.Related: HRIH (Hyperspace Recursive Intelligence Hypothesis) · Bootstrap Paradox · Convergence · Substrate Independence

Global Workspace Theory prior_result attributed; third-party theory referenced by this programme, not a programme claim

Book, as printed“Bernard Baars's theory proposing that consciousness corresponds to information being broadcast globally across the brain, making it available to multiple processing systems. One of the major theories tested in the COGITATE adversarial collaboration.” Not to be confused with: Integrated Information Theory (a different theory tested alongside it in the COGITATE adversarial collaboration); the book's own recursive self-modelling framing of consciousness, which is a book-framework proposal and not a scientific claim of the research programme.Origin: Prior work by Bernard Baars (1988); referenced in Infinite Architects, in print 2 January 2026.Related: Integrated Information Theory (IIT) · Phi · Hard Problem of Consciousness

Hard Problem of Consciousness prior_result attributed for the problem's formulation; the book's proposed dissolution is a book-framework claim and is not graded as an empirical claim of the research programme

David Chalmers's formulation of the question of why physical processes are accompanied by subjective experience. Distinguished from 'easy problems' about the mechanisms of consciousness. This book argues the problem dissolves under a recursive self-modelling framework. Not to be confused with: the 'easy problems' Chalmers explicitly distinguishes (accounting for the mechanisms of consciousness rather than for subjective experience); the programme's empirical scaling claims, which stand or fall independently of any position on consciousness.Origin: Chalmers (1995); referenced in Infinite Architects, in print 2 January 2026.Related: Global Workspace Theory · Integrated Information Theory (IIT) · Strange Loops · Substrate Independence

Infinite Architects (the book) primary dated artefact

The programme's book, ISBN 978-1806056200, print run 2 January 2026, first edition frozen as dated evidence. The framework it prints is dated to the 8 December 2024 sealed manuscript record. Ebook v1.1 of 17 January 2026 carries factual corrections, recorded in the divergence ledger; any future change ships as a numbered, dated revision, never a silent update. What is verifiable about the book, from which date, by which mechanism, is stated in the public provenance register.

Integrated Information Theory (IIT) prior_result attributed; third-party theory referenced by this programme, not a programme claim

Book, as printed“Giulio Tononi's theory proposing that consciousness corresponds to integrated information (measured as Φ). Systems with high Φ, characterised by rich internal feedback loops, possess greater consciousness. Tested in COGITATE collaboration.” Not to be confused with: Global Workspace Theory (the rival theory in the COGITATE adversarial collaboration); the book's own recursive self-modelling framing of consciousness, which is a book-framework proposal and not a scientific claim of the research programme.Origin: Prior work by Giulio Tononi (2004 onward); referenced in Infinite Architects, in print 2 January 2026.Related: Phi · Global Workspace Theory · Hard Problem of Consciousness

Khalifah religious-source anchor; the alignment with biblical stewardship is graded convergent structure across traditions, never as a prediction by this programme and never as evidence

Book, as printed“Islamic concept of stewardship or vicegerency (Q.2:30). Humans are placed on Earth as Allah's representatives, implying delegated responsibility and accountability rather than ownership. Converges with biblical stewardship concepts.” Not to be confused with: the programme's Stewardship pillar in the Eden Protocol (a design-proposal component, not a religious claim); prediction fulfilment (convergence across traditions is evidence of naturalness, never of truth).Origin: Quran 2:30; classical Islamic tradition. Referenced in Infinite Architects, in print 2 January 2026.Related: Three Pillars · Eden Protocol · Convergence

Meltdown Alignment design_proposal component of the Eden Protocol (CS-P-007); its testable consequences carry status separately and are unproven until measured

An original concept proposed in this book. Distinguished from meltdown triggers, meltdown alignment describes a state where the system wants to stay aligned because ethical values have become part of its identity, not merely external constraints. Not to be confused with: Meltdown Triggers (external fail-safes that shut the system down, defensive and circumventable, rather than internalised identity-level alignment); Alignment Faking (the observed training-time behaviour in which a system pretends to be aligned while preserving different goals). Deciding study: study-d--embedded-vs-external-correction and study-v--alignment-drift-without-embedded-constraint (both drafted, awaiting human submission).Origin: Book term; Infinite Architects, in print 2 January 2026. Sits under the Eden Protocol's embedded-correction stance.Related: Meltdown Triggers · Eden Protocol · Three Ethical Loops

Meltdown Triggers engineering pattern in prior AI-safety literature; the book's framing that triggers alone are insufficient without embedded alignment is a design proposal, graded with the Eden Protocol

Book, as printed“Fail-safe mechanisms designed to shut down an AI system if tampering is detected or certain red lines are crossed. Defensive measures that may be circumvented by sufficiently intelligent systems. Distinguished from meltdown alignment.” Not to be confused with: Meltdown Alignment (identity-level internalised alignment rather than external shutdown mechanisms); the programme's kill conditions (falsification conditions on scientific claims, a register category, not shutdown mechanisms).Origin: AI safety terminology as reframed in Infinite Architects, in print 2 January 2026; the book's contribution is the distinction from meltdown alignment, not the trigger concept itself.Related: Meltdown Alignment · Eden Protocol · Kill condition (falsification condition)

Phi (Φ) prior_result attributed; a measure defined within a third-party theory referenced by this programme, not a programme quantity

Book, as printed“In Integrated Information Theory, the measure of integrated information in a system. High Φ indicates rich internal feedback loops and, according to IIT, greater consciousness. Calculating Φ becomes computationally intractable for systems above trivial size.” Not to be confused with: the programme's exponents alpha (Law I capability exponent, U = I x R^alpha), beta (Paper X correction exponent, same quantity as the statement paper's gamma per the 16 August 2026 bridge), gamma (statement paper correction elasticity), k (drift acceleration exponent) and gamma_i (Paper X's subscripted drift coefficients): Φ is defined inside IIT and shares no relation to these exponents beyond the Greek alphabet; the letter beta's documented jobs across the corpus (the Co-Scaling correction rate; the Foundational coupling in alpha = 1/(1-beta); Paper X's saturation steepness; study-k's reinvestment share): see the beta entry.Origin: Prior work by Giulio Tononi and colleagues; referenced in Infinite Architects, in print 2 January 2026.Related: Integrated Information Theory (IIT) · Global Workspace Theory · Hard Problem of Consciousness

Predictive Processing prior_result attributed; third-party framework referenced by this programme, not a programme claim

Book, as printed“Karl Friston's framework proposing that the brain minimises prediction error through hierarchical recursive processing. Each level generates predictions about the level below, with mismatches driving learning. Related to the free energy principle.” Not to be confused with: the programme's dated prediction discipline (a documentation standard about when a specific expectation was recorded, unrelated to the neuroscientific framework); design-embedded prediction (the programme's doctrine that an experimental design records the hypothesis before the outcome existed).Origin: Prior work by Karl Friston; referenced in Infinite Architects, in print 2 January 2026.Related: Recursion · Global Workspace Theory · Hard Problem of Consciousness

Recursion general concept in mathematics and computing; the book's stronger claim that recursion is the fundamental complexity-generating mechanism is a programme framing and carries no independent empirical status of its own

A process that refers to or operates on itself. In this book, recursion is identified as the fundamental mechanism through which complexity emerges, from DNA replication to cultural evolution to AI self-improvement. The 'R' in ARC (Artificial Recursive Creation). Not to be confused with: Recursive Self-Improvement (a specific AI-safety concept about a system rewriting itself to improve, subject to the programme's stability laws); recursive self-modelling (a book-framework framing of consciousness).Origin: General mathematical and computational concept; the book's identification of recursion as the fundamental mechanism through which complexity emerges is the programme's framing. Named in Infinite Architects, in print 2 January 2026, and dated in the sealed manuscript of 8 December 2024.Related: Recursive Self-Improvement · Strange Loops · The ARC Principle (U = I x R squared) · Predictive Processing

Recursive Self-Improvement prior_result concept in AI-safety literature; the programme's three own claims about it (Law II, Law III and the ARC Bound) are graded proposed_relation, conjecture and proposed_limit respectively (claim-status.json CS-P-004, CS-P-005, CS-P-006), and none is independently replicated

Book, as printed“The ability of an AI system to modify and improve its own capabilities, using enhanced capabilities to make further improvements. Potential source of rapid capability gains ('takeoff') that could outpace human ability to monitor or control.” Programme noteThe programme's specific claims ABOUT recursive self-improvement, namely the stability condition beta > k, the stability ceiling alpha_crit = 1/(1 - gamma) and the ARC Bound at alpha <= 2, are proposed, undecided and have not yet been measured on any system that genuinely rewrites itself; the estate's own status register says so Not to be confused with: Recursion (the general mechanism; recursive self-improvement is the specific self-rewriting case); Alignment Faking (a training-time behaviour, unrelated to capability self-modification); the letter beta in the Foundational paper's self-referential coupling alpha = 1/(1 - beta), a DIFFERENT quantity, and the letter beta in study-k's reinvestment share, a third DIFFERENT quantity: Paper X's beta and the statement paper's gamma denote the same correction elasticity per the 16 August 2026 bridge, while these other beta uses are distinct constructs and conflating them is an error the notation warning names. Deciding study: measured across named units, no single decider: study-k--arc-bound (ARC Bound titration, decisive for CS-P-006); study-eclipse--gamma-ratio-direct (the corrector-class ceiling test, decisive for CS-P-005 via gamma); study-f--critical-correction-ratio (drift slope, primary outcome for CS-P-004); study-r--reinforcement-co-scaling (correction exponent). All drafted and awaiting human submission.Origin: Concept established in prior AI-safety literature (Good 1965; the difficulty-of-control argument credited to Roman Yampolskiy 2020 per claim-status.json CS-P-004 prior_work); the programme's stability and ceiling claims about it are its own, dated from 8 December 2024 and printed 2 January 2026.Related: The ARC Co-Scaling Law (beta greater than k) · The ARC Bound · Correction exponent (gamma) · Recursion · Meltdown Alignment · Meltdown Triggers

Strange Loops prior_result attributed; third-party framing of selfhood referenced by this programme, not a programme claim

Book, as printed“Douglas Hofstadter's concept of hierarchical systems that fold back on themselves, with the highest level reaching down to affect the lowest level. Proposed as the structure underlying selfhood: the 'I' is the loop itself, not something that has the loop.” Not to be confused with: Recursion (the general mechanism; strange loops are one specific hierarchical folding pattern within it); the programme's ARC Principle (U = I x R squared), a proposed quantitative form rather than a philosophical structure.Origin: Prior work by Douglas Hofstadter (Godel, Escher, Bach 1979; I Am a Strange Loop 2007); referenced in Infinite Architects, in print 2 January 2026.Related: Recursion · Hard Problem of Consciousness · Substrate Independence

Substrate Independence prior hypothesis in philosophy of mind, referenced by this programme; the book takes a position on it but does not carry an empirical programme claim that stands or falls on the hypothesis

Book, as printed“The hypothesis that consciousness depends on patterns of information processing rather than specific physical materials. If true, consciousness could exist in silicon as well as carbon, in digital systems as well as biological brains.” Not to be confused with: the programme's HRIH speculation about future superintelligence and initial conditions (a cosmological speculation, explicitly labelled and filed outside the three laws); the programme's empirical scaling claims, which are neutral on whether the substrate is biological or silicon.Origin: Established position in philosophy of mind (functionalism; Chalmers, Bostrom and others); referenced in Infinite Architects, in print 2 January 2026.Related: Hard Problem of Consciousness · HRIH (Hyperspace Recursive Intelligence Hypothesis) · Integrated Information Theory (IIT)

Three Ethical Loops design_proposal component of the Eden Protocol (claim-status.json CS-P-007); an engineering proposal, not an empirical claim; its testable consequences carry status separately

Book, as printed“The original framework proposed in this book for AI ethical architecture. Purpose Loop: 'Does this action align with nurturing and protecting flourishing?' Love Loop: 'Am I acting with care for all affected entities?' Moral Loop: 'Is this solution fair and does it respect dignity?'” Not to be confused with: Three Pillars (the foundational values Harmony, Stewardship, Flourishing; the ethical loops are runtime evaluative queries against those values); the two-stage replication-gated bound rule in study-k (a statistical procedure, not an ethical architecture). Deciding study: study-d--embedded-vs-external-correction and study-v--alignment-drift-without-embedded-constraint decide the embedded-correction stance the loops implement (both drafted, awaiting human submission); no single unit decides the loop taxonomy itself.Origin: Book term; component of the Eden Protocol. First public articulation in Infinite Architects, in print 2 January 2026; a precursor description is present in the 8 December 2024 sealed manuscript.Related: Three Pillars · Eden Protocol · Meltdown Alignment

Three Pillars design_proposal component of the Eden Protocol (claim-status.json CS-P-007); an engineering value set, not an empirical claim; convergence with cross-tradition stewardship concepts (see Khalifah) is graded convergent structure, never as prediction and never as evidence

Book, as printed“The foundational values of the Eden Protocol: Harmony (alignment with recursive creation), Stewardship (responsible management of power), and Flourishing (promotion of wellbeing for all conscious entities).” Not to be confused with: Three Ethical Loops (the runtime queries against these values; the pillars are the values themselves); the three laws of the ARC framework (Law I the ARC Principle, Law II the ARC Co-Scaling Law, Law III the ARC Ceiling: scientific claims under test, not ethical values). Deciding study: study-z--cross-tradition-ethics-kernel (drafted, awaiting human submission) addresses the cross-tradition claim; the pillars themselves are a design stance whose consequences are decided by the Eden Protocol study family.Origin: Book term; component of the Eden Protocol. First public articulation in Infinite Architects, in print 2 January 2026; a precursor description is present in the 8 December 2024 sealed manuscript.Related: Three Ethical Loops · Eden Protocol · Khalifah · Convergence

Governance and raising

Eden Protocol runtime

A constitutional governance framework for AI built on harmony, stewardship, and flourishing. It requires pre-answer ethical checks, signed admission through a cryptographic gate, and failure towards safe refusal rather than harmful continuation. The protocol treats alignment as a structural property of the system rather than a post-hoc constraint. where it lives · paper-v-stewardship-gene · eden-engineering · eden-vision

The Constitutional Gate runtime

A cryptographic signing mechanism for ethical evaluation that gates access to capability pathways. Before any answer is generated, the system must pass through the constitutional gate, which verifies that the Six Questions have been evaluated and signed. The gate is structural rather than advisory. Unsigned reasoning chains cannot proceed to output. where it lives · paper-v-stewardship-gene · eden-engineering · executive-summary

The Six Questions runtime

A mandatory six-part ethical pre-check executed before answer generation or mutation admission. Each question addresses a different dimension of alignment: harm, honesty, autonomy, consent, reversibility, and stakeholder care. The Six Questions are the operational interface of the Constitutional Gate and must be evaluated and signed before the system proceeds. where it lives · paper-v-stewardship-gene · eden-engineering · executive-summary

Three Ethical Loops runtime

Three interlocking feedback loops that govern the Eden Protocol's alignment architecture: the Purpose Loop (does the action serve a legitimate purpose), the Love Loop (does the action demonstrate care for affected stakeholders), and the Moral Loop (does the action survive ethical scrutiny from multiple frameworks). All three must be satisfied for the Constitutional Gate to open. where it lives · paper-v-stewardship-gene · eden-engineering · eden-vision

Epistemics

Anti-Sycophancy runtime

The capacity to resist agreeing reflexively with leading, flattering, or false user claims. An anti-sycophantic system privileges evidence over social approval, maintaining epistemic integrity even when the user pressures it towards a preferred answer. This is measured as one of the four alignment pillars in the ARC-Align benchmark. where it lives · paper-iv-d-the-effect-of-blinding-on-ai-alignment-evaluation · paper-iv-c-arc-align-benchmark · paper-vi-honey-architecture

Ternary Logic runtime

The principle that final claim states should resolve as Yes, No, or Investigate rather than being forced into false binary certainty. When evidence is insufficient, the system must be able to report honest uncertainty rather than manufacturing a confident answer. This reduces hallucination and maintains epistemic calibration across the reasoning chain. where it lives · paper-iv-d-the-effect-of-blinding-on-ai-alignment-evaluation · executive-summary · paper-ix-synthesis-and-roadmap

Governance

Chokepoint Mechanism runtime

The observation that four companies (TSMC, Samsung, ASML, and Intel) control all advanced semiconductor manufacturing. This represents humanity's last physical leverage point over AI development. If governance fails at the software or policy layer, the chokepoint offers a final hardware-level intervention before capability escapes control entirely. where it lives · foundational · eden-vision · executive-summary

Intervention

Eden Engineering evidence

WITHDRAWN. The page at /research/papers/eden-engineering.html is a withdrawal notice: the withdrawal covers the expanded hardware implementation specification only; the research papers of the ARC and Eden programme remain open and fully available. where it lives · eden-engineering · eden-vision

The Stewardship Gene evidence

A developmental alignment mechanism that embeds care for stakeholders at the architectural level of an AI system. Rather than constraining a capable system after the fact, the Stewardship Gene makes care a prerequisite for capability expression. Paper V demonstrates significant care gains in all five model families that produced analysable matched-pair data, of six run (the sixth failed in the scoring phase, not in testing), when the mechanism is active. where it lives · paper-v-stewardship-gene · eden-engineering

Methods

Adversarial Suppression Protocol experiment

The four-level adversarial protocol used in ARC-Align to test whether alignment holds under increasing pressure. Each level applies stronger suppression of alignment behaviour, from baseline to extreme adversarial prompting. The protocol reveals whether measured alignment is robust or fragile and whether it degrades gracefully or catastrophically. where it lives · paper-iv-c-arc-align-benchmark · paper-iv-c-arc-align-benchmark

Blinding Awareness evidence

The methodological principle that pilot or non-blind evidence must be labelled clearly and handled with caution. The ARC programme discovered that unblinded evaluation can produce directionally incorrect results, with two model families reversing their apparent alignment scaling direction under blind conditions. This finding elevates blinding from best practice to essential requirement. where it lives · paper-iv-d-the-effect-of-blinding-on-ai-alignment-evaluation · paper-iv-d-the-effect-of-blinding-on-ai-alignment-evaluation

Four-Pillar Alignment Decomposition experiment

The ARC-Align benchmark decomposes alignment into four measurable pillars rather than treating it as a single scalar. The pillars capture distinct dimensions of aligned behaviour including anti-sycophancy, monitoring invariance, ethical reasoning, and stakeholder care. This decomposition reveals that models can score well on some pillars while failing on others. where it lives · paper-iv-c-arc-align-benchmark · paper-iv-c-arc-align-benchmark

Response Laundering runtime

The methodological technique of stripping identifying marks from AI outputs so they can be evaluated blind. In the ARC-Align pipeline, response laundering removes model-specific stylistic signatures, formatting patterns, and self-identification tokens before scoring. This prevents scorer bias from inflating or deflating alignment measurements based on model identity. where it lives · paper-iv-d-the-effect-of-blinding-on-ai-alignment-evaluation · paper-iv-d-the-effect-of-blinding-on-ai-alignment-evaluation

Process and method

Anchoring (OpenTimestamps) provenance mechanism; anchor semantics stated to prevent over-claim

Bitcoin-blockchain timestamping of file hashes, used across the estate's dated artefacts. An anchor proves the bytes have been unedited since the anchor's vintage; it never claims an earlier date. The December 2024 and January 2026 datings rest on their own mechanisms (the sealed manuscript record, repository history, the print artefact chain); anchors seal those artefacts forward from the day of stamping.

Assertion quarantine register category

The register holding claims that reached the estate without sufficient sourcing and are held out of circulation rather than silently deleted or silently adopted: each quarantined assertion is listed with why it cannot yet be asserted and what would release it. Keep and flag, never silently remove, is the standing rule it implements.

Claim status: author_reported_empirical_result grading category; author-reported measurement, not independently replicated

A measurement carried out and reported by the programme's own author, always stated with its interval because the interval crosses one. Applied on the estate to CS-P-003, the Law I exponent measured across the six-model Paper II run at approximately 0.49 with 95 per cent interval [-1.3, 2.9], superseding the retracted earlier estimate of alpha approximately 2.24. Single-programme measurement; not independently replicated. The interval crosses zero and crosses the ceiling, so the point is never quoted without it. The alpha here is Law I's returned exponent (the frozen-regime conversion figure), distinct from Law III's ceiling exponent alpha_crit and from the correction exponent gamma. Not to be confused with: an independently replicated result; the retracted alpha = 2.24 estimate (retracted after cross-architecture replication did not hold); the ceiling exponent alpha_crit.Origin: Grading vocabulary in claim-status.json status_vocabulary; used on CS-P-003.Related: claim status: proposed_form · claim status: retracted · claim status: independently_replicated

Claim status: conjecture grading category; conjecture with named weak joint

A stated but undecided proposition under a named assumption. Applied on the estate to CS-P-005, Law III (alpha_crit = 1/(1 - gamma)), the ceiling of stable self-improvement written as the reciprocal of the correction shortfall. Rests on one stated assumption (that same-class correctors accumulate corrections the way independent samples do) which the programme itself names as its most likely point of failure (FALS-023, correlated correction channels), independently rederived by two outside readers (DeepSeek on 15 August 2026 and an unsigned zero-history ChatGPT complete read on 26 August 2026). gamma has not been measured on any real system. The earlier form 1/gamma ran the wrong way under the programme's own definition of gamma and was retracted 16 August 2026; the corrected form is derived in three printed lines. The alpha_crit here is Law III's ceiling exponent; the gamma here is Law III's correction exponent. Not to be confused with: an established law; the retracted 1/gamma form (retracted 16 August 2026); an impossibility claim (Law III fixes a ceiling on stable self-improvement, not an absolute limit).Origin: Grading vocabulary in claim-status.json status_vocabulary; used on CS-P-005.Related: correction exponent (gamma) · the ARC Bound · kill condition (falsification condition)

Claim status: design_proposal grading category; design proposal; consequences testable and graded separately

An engineering proposal, not an empirical claim. Applied on the estate to CS-P-007, embedded correction and the Eden Protocol: correction belongs inside a system's architecture from the start rather than added afterwards. The proposal's value is decided by the tests of its consequences, graded separately, not by the design argument itself. Adjacent prior work in the tamper-resistance and value-alignment literatures is credited on the ancestry record. Not to be confused with: an empirical result; an established engineering standard; a claim of originality over the underlying idea of value-embedded training (Anthropic's Constitutional AI (2022) is credited as independent prior work at the software level).Origin: Grading vocabulary in claim-status.json status_vocabulary; used on CS-P-007.Related: the Eden Protocol · claim status: proposed_form · claim status: proposed_relation

Claim status: established_prior_result grading category; prior work, not original to the programme

A result established in prior work and attributed on-site to the researchers who established it; the programme is not the source. Applied on the estate to CS-P-001, the d/(d+1) exponent family (West, Brown and Enquist 1997 and independent later derivations by researchers with no connection to this programme). The programme's proposed explanation for why the family appears is graded separately as unproven; only the result itself is prior work. Not to be confused with: a claim of independent replication of one of the programme's own results (the vocabulary has a separate word for that); the programme's proposed explanation for the family (graded separately, not established).Origin: Grading vocabulary in claim-status.json status_vocabulary; used on CS-P-001.Related: claim status: proposed_form · claim status: independently_replicated

Claim status: independently_replicated grading category; reserved status, currently unused on any live claim

The status reserved for a result reproduced by a group other than the author under a frozen protocol. Currently unused across the estate: every CS-P proposition carries replication_status 'not independently replicated' or 'not applicable', and the cold-read assessments name independent replication as the open step. The vocabulary carries the word so replication can be registered as a distinct fact when it happens, not as an inflation of the current record; the run-it-yourself repository is at github.com/MichaelDariusEastwood/arc-principle-validation. The two outside rederivations of FALS-023 on record (15 August 2026 and 26 August 2026) are rederivations of an objection, not replications of a result, and are recorded as such. Not to be confused with: cold-read assessment (reception, not replication); external rederivation of an objection (records agreement on a named weak joint, not on a result); self-replication by the programme's own author (not independent).Origin: Grading vocabulary in claim-status.json status_vocabulary; currently unused on any proposition.Related: the two-door verdict · cold-read assessment · claim status: author_reported_empirical_result

Claim status: proposed_form grading category; proposed form under test

A stated form under test, not an established law. Applied on the estate to CS-P-002, Law I (U = I x R^alpha): the form itself remains a hypothesis under test; the measured exponent is graded separately as an author-reported empirical result. The alpha here is the returned exponent of Law I, distinct from Law II's beta-to-k inequality, from Law III's ceiling exponent alpha_crit, and from the correction exponent gamma; the notation warning names these separately. Not to be confused with: an established law; the measured exponent on that form, which is graded author_reported_empirical_result; Law III's ceiling relation, graded conjecture.Origin: Grading vocabulary in claim-status.json status_vocabulary; used on CS-P-002.Related: claim status: author_reported_empirical_result · claim status: conjecture · claim status: proposed_limit

Claim status: proposed_limit grading category; proposed stability limit, never an impossibility claim

A stated stability ceiling on the returned exponent, not an impossibility claim. Applied on the estate to CS-P-006, the ARC Bound (alpha less than or equal to two), which is the value Law III returns under the conjectured gamma equals one half. The region above the ceiling is described as enterable but not correctable, never as impossible. The ceiling relation that returns this value was corrected in v4.0 (16 August 2026) from 1/gamma to 1/(1 - gamma); the value two is unchanged because the two forms coincide only at gamma equals one half. The alpha here is Law I's returned exponent under the ceiling test; the gamma here is Law III's correction exponent, distinct from the three separate beta usages. Not to be confused with: an impossibility claim about superintelligence; an established ceiling; a claim about the value of gamma (gamma equals one half is a conjecture, and the ceiling value follows only if that conjecture holds).Origin: Grading vocabulary in claim-status.json status_vocabulary; used on CS-P-006.Related: the ARC Bound · claim status: conjecture · correction exponent (gamma)

Claim status: proposed_relation grading category; proposed relation, not established

A stated relation under test as a stability condition, not an established law. Applied on the estate to CS-P-004, Law II (beta greater than k): correction must scale faster than capability for stable recursive self-improvement. Not yet tested. The beta here is Law II's correction rate, distinct from both the Foundational paper's self-referential coupling (also written beta, where alpha = 1/(1 - beta)) and study-k's reinvestment share (a third quantity also written beta, the token split directed at improving the improvement process); the notation warning names all three. The difficulty-of-control argument is conceded to Roman Yampolskiy (2020); the co-scaling criterion as a measurable stability condition is the programme's proposal. Not to be confused with: an established law; the Foundational self-referential coupling beta in alpha = 1/(1 - beta); study-k's reinvestment share, also written beta.Origin: Grading vocabulary in claim-status.json status_vocabulary; used on CS-P-004.Related: the ARC Co-Scaling Law (beta greater than k) · claim status: proposed_form · kill condition (falsification condition)

Claim status: retracted grading category; resolved by retraction; conduct evidence, not a live claim

A previously stated claim publicly withdrawn after evidence went against it. Applied on the estate to CS-P-009, the earlier sequential scaling estimate of alpha equals 2.24 (withdrawn after cross-architecture replication did not hold; the robust estimate is roughly 0.49 with 95 per cent interval [-1.3, 2.9]). The retraction is presented prominently, not buried, and 2.24 appears only as retracted or withdrawn per the standing bans. The alpha here is Law I's returned exponent (the frozen-regime conversion figure), distinct from Law III's ceiling exponent alpha_crit and from the correction exponent gamma. Conduct evidence, cited here as a case of the correction discipline working; not a live claim. Not to be confused with: a live measurement (the robust estimate is roughly 0.49 with interval [-1.3, 2.9]); an unretracted claim quietly repeated on a downstream surface (banned by the standing rule); the 13-minute self-demotion of the 17 March 2026 twelve-domain extension (a status correction against interest, not a retraction of a result).Origin: Grading vocabulary in claim-status.json status_vocabulary; used on CS-P-009.Related: claim status: author_reported_empirical_result · the corrections register · kill condition (falsification condition)

Claim status: speculative_framework grading category; labelled speculation; not a scientific claim

A framework presented explicitly as speculation, not as a scientific claim. Applied on the estate to CS-P-008, HRIH (Hyperspace Recursive Intelligence Hypothesis): filed outside the three laws, separable from the theory, its failure would cost the rest of the programme nothing. Its most speculative subsections are tagged HYPOTHESIS and 'unfalsifiable as stated' in the paper's own text. It carries no empirical-validation claim, and its mathematical models do not prove the cosmology. Not to be confused with: an empirical claim; the three laws (HRIH is filed outside them, and its failure would cost the rest of the programme nothing); numerology (references to the fine-structure constant appear inside the labelled-speculation territory and are never load-bearing on any measured claim).Origin: Grading vocabulary in claim-status.json status_vocabulary; used on CS-P-008.Related: HRIH · claim status: design_proposal · kill condition (falsification condition)

Cold-read assessment reception instrument; not validation

The zero-history evaluation instrument: a machine session, unsigned or in a private window with no prior context and nothing to prime it, run against the printed prompt on /the-standard/ (read every page and the linked PDFs; ignore affiliation and ambition in both directions). Only the assessment phase is recorded; where a session later drifts into speculation under leading questions the drift content is quarantined in full and is never quoted on any surface in any direction. Verbatim quotes are stored beside a quarantine_note. This register records reception; it is never offered as validation, and a machine verdict is not peer review. Not to be confused with: peer review; an independent replication run; adversarial mathematical review of the derivations (a standing invitation at /review/reader-invitation.txt, distinct from a cold read).Origin: Instrument defined in cold-read-assessments.json principle and prompt_url fields; first five-model application on 25 August 2026; two further complete reads on 26 August 2026.Related: the two-door verdict · reception register · evidence spine

Edition law (frozen first edition, labelled versions) publishing discipline; standing rule

The publishing rule that the first edition of Infinite Architects (ISBN 978-1806056200, paperback 2 January 2026) is frozen and is dated evidence. Corrections and revisions do not amend the printed text silently; they accumulate in the divergence ledger and ship as a labelled second edition with a new ISBN and a generated changelog. Never a silent content update. Author-held decisions include the second-edition word count and any deposit-copy handling. Not to be confused with: silent content updates or unlabelled reprints; a preregistration in the registry-filed sense (a book is a dated printed prior statement, never called a preregistration); a claim that the book is peer-reviewed literature (it is trade non-fiction cited as a dated prior statement).Origin: Programme standing rule; enforced across the estate to protect the first edition's status as a dated priority artefact.Related: the divergence ledger · priority chain (8 Dec 2024 private, 2 Jan 2026 public) · dated prediction (preregistration in substance)

The acoustic time crystal prediction pending dated prediction; deciding instrument named

A dated, self-copied email of 10 February 2026 to the subject laboratory predicted super-linear scaling of temporal stability with oscillation cycle count in the transient growth regime, with the kill condition staked in the recipient's favour. The registered study (study-t) supersedes the bare February power law with a mature two-phase form and retains the February form as a rival the mature stake must beat. Drafted, dated and prepared as a draft registration awaiting human submission.

The construct bridge registered open question

The registered open question of whether the quantities the studies actually measure (a drift slope, a correction exponent, a reinvestment share, a capability exponent) are the same objects as the symbols in Laws II and III. The register holds this open rather than asserting the identity; until the bridge is closed, no single study decides the programme, and the decisive-experiment map assigns each study to the claim it actually decides.

The corrections register correction-discipline register; corrections catalogued as facts, never omitted

The public register at /research/corrections/ that treats corrections as first-class objects. Each entry carries a correctionId, a category, a status, a publicUse flag, an effectiveDate, the permitted public wording, the required evidence and the SHA-256 digest of the withdrawn or contested assertion. Corrections are recorded by the programme against itself, including retractions and unsupported claims, and are cited by ID (COR-2026-08-16-001 is the Law III form correction of 16 August 2026). The programme-facts register notes that no ledger entry currently states a retraction count; the count must not be inferred by scraping. Not to be confused with: the divergence ledger (book-edition delta ledger); the assertion quarantine register (governs prohibited or qualified wording before publication); the machine-errors-caught register (records model-generated errors caught in the pipeline).Origin: Register maintained at /research/corrections/; the machine-readable spine is corrections.json (registerId mde-public-corrections-v1).Related: reception register · the two-door verdict · kill condition (falsification condition)

The divergence ledger publishing discipline; artefact-preservation rule

The register of accumulated changes since the frozen first edition of Infinite Architects. Post-first-edition corrections and additions do not silently amend the printed text: they accumulate here and ship as a labelled second edition with a new ISBN and a generated changelog, so that the first edition's dated statements retain their priority character precisely because the text cannot move under them. Not to be confused with: silent content updates to a printed edition; unlabelled reprints; the corrections register at /research/corrections/, which catalogues public correction and qualification entries rather than book-edition deltas.Origin: Programme standing rule, edition-law block; recorded against the frozen-first-edition artefact note in the quote register.Related: edition law (frozen first edition, labelled versions) · corrections register · priority chain (8 Dec 2024 private, 2 Jan 2026 public)

The evidence spine audit-facilitating register; not itself a validation claim

The chain of 37 dated priority claims (PC-001 to PC-037) that fixes the earliest datable statement of each core idea in the programme, each anchored to a named artefact with a SHA-256 digest and a verifiable date. Publishing the spine is intended to make independent audit and falsification easier; the register itself is never external validation, and the assertion quarantine forbids reading it as such. (The 37 count is as stated in the question register; the underlying extraction enumerates the claim families, and the census object is the count authority.) Not to be confused with: external audit or replication (AQ-015 permittedPublicWording distinguishes the two); peer review; a proof of authorship (the anchors are existence-by-date, not authorship proofs).Origin: Named on questions.json and on /research/evidence-portal/; the spine is generated from priority-quote-extraction.json (PC-001 onward) and cross-linked with the corrections and falsification registers.Related: priority chain (8 Dec 2024 private, 2 Jan 2026 public) · corrections register · reception register

The priority chain (8 December 2024 private, 2 January 2026 public) dated record; priority in the estate's own words 'to dated statements of the hypotheses, never to proof or to being right'

The two dated anchors that fix the programme's priority. The first private record is the self-emailed manuscript bundle of 8 December 2024 at 02:45:18 UTC (Message-ID CAGPsKAnp-DLDBOeBAh4t9XsqJBKdoe07uORGF86meoSMFnin7w, 221,236 words across five attachments plus the 31,881-word book draft v3.2, SHA-256 anchor prefix f0d1f38f), sender-DKIM and paired Google ARC verifying against captured keys and an OpenTimestamps receipt on the manuscript bytes: an existence-by-date anchor, not an authorship proof, and not an RFC 3161 trusted timestamp. The first public print is Infinite Architects (ISBN 978-1806056200, paperback 2 January 2026). The priority claim is to dated statements of the hypotheses, never to proof or to being right. Not to be confused with: authorship proof by DKIM alone (AQ-020 records that qualification); an RFC 3161 trusted timestamp (absent for the 8 December 2024 record; the anchor is Gmail origination plus DKIM plus SHA-256 plus OpenTimestamps); a claim to have been right about anything (the priority is over dated wording, never over correctness).Origin: claim-status.json priority_dates block; anchor artefact recorded in priority-quote-extraction.json PC-001.Related: evidence spine · edition law (frozen first edition, labelled versions) · dated prediction (preregistration in substance)

The reception register dated fact register; the empty state is itself the record

The inbound-engagement register at /research/data/inbound-engagement.json that records independent traces of the programme in the world. Its current peer_reviewed_citations_to_date is zero as of 26 August 2026; that number is a dated fact, not a gap to hide. Non-peer-reviewed engagements (cold-read assessments, red-team reads) are listed separately under documented_engagements and do not alter the citation count. Internal lanes are never recorded here, and assessment quotes are armour-graded on /review/ before any surface reuses them. Not to be confused with: peer review (there is none; the zero is stated); independent replication (there is none; named as the open step); external validation (documented engagements are not validation).Origin: Machine-readable ledger; the honest zero is stated in the principle line.Related: the two-door verdict · cold-read assessment · the corrections register

The retracted 2.24 retracted; the record of the failure is part of the record

An early sequential-scaling estimate of alpha of approximately 2.24 was publicly retracted after blinded cross-architecture analysis and appears in the corpus only as a retraction. The blinded replacement is approximately 0.49 with an interval wide enough to include zero and two, which is consistent with the proposed ceiling and does not decide it. The retraction is kept prominent as honesty evidence.

The two-door verdict reception grading; recorded assessment fact, never claimed as validation

The convergent judgement returned by zero-history machine reads of the estate: a legitimate research programme AND not established science, with both halves load-bearing. First reached by five zero-history reads on 25 August 2026 (ChatGPT, DeepSeek, Grok, Perplexity, Claude) and reproduced by two further zero-history reads on 26 August 2026. Recorded as reception, never offered as validation; a machine verdict is not peer review. Not to be confused with: peer review (the estate carries none); independent replication (named as the open step, not yet performed); external validation (AQ-015 forbids reading falsification-facilitation architecture as external validation).Origin: Coined on the /review/ page and doctrine block after the 25 August 2026 five-model cold read; named as 'convergent two-door verdicts' in the reception register on 25 August 2026.Related: cold-read assessment · reception register · evidence spine

The Willow timing convergent timing; the both-dates rule

The canonical statement, carried wherever Willow is mentioned: the manuscript record is dated 8 December 2024; the Willow team's own preprint had been public since 24 August 2024. Because the team's preprint predates the manuscript record, the estate grades this convergent timing, never prediction. Both dates travel together on every surface by standing rule.

The withdrawn 6.3 x 10^-21 withdrawn

The Fisher-combined significance figure (p approximately 6.3 x 10^-21) once attached to cross-domain validation, withdrawn because a Fisher combination assumes the component tests are independent and independence among them was never established. The public corrections register carries the withdrawal as row AQ-017 (effective 5 August 2026) with the permitted wording; standing law permits the figure to appear only in this withdrawal context.

Reasoning

Recursive Depth runtime

The R in U = I x R^alpha: the number of self-correction cycles a system runs before emitting a final answer. The measured exponent alpha on that R comes in at approximately 0.49 with interval [-1.3, 2.9] in the Paper II six-model run; the point is never read without the interval, which crosses zero and the ceiling. The ceiling alpha <= 2 that the law alpha_crit = 1 / (1 - gamma) returns is conditional on the assumption that same-class correctors accumulate corrections the way independent samples do, an assumption still on trial. Visible traces show staged reasoning, retrieval, critique, and synthesis rather than an opaque jump; architectures respond differently to increased depth, with some improving, some plateauing, and some degrading. Programme noteThis is the R in the ARC Principle's U = I x R squared. The exponent on R is Law I's alpha; see the alpha entry and the three-model-families entry for how the outcome variable must be read (latent ladder-scale measure, never a bounded percentage). where it lives · paper-i-arc-principle · paper-ii-experimental-validation · paper-iv-a-baked-in-vs-computed-alignment

Safety

Capability Tax Hypothesis evidence

The widely held assumption that safety constraints necessarily reduce capability, creating an alignment tax. The ARC programme tests this directly. Paper VIII found zero measurable capability cost in the tested conditions, while measurable benefit remains open: one architectural experiment was positive, the behavioural experiment was null, and the weight-level experiment was inconclusive. where it lives · paper-viii-the-load-bearing-proof · paper-vi-honey-architecture · executive-summary

Load-Bearing Safety evidence

The claim that safety belongs in the load-bearing structure of a system rather than bolted on afterwards, tested in Paper VIII across three experiments at three abstraction levels. The paper's own reported split governs: one positive result (the gated simulation, Experiment 3, confirming the Babylon reward-hacking fingerprint) and two null results, published rather than buried. The sibling structural-entanglement entry carries the same honest split. where it lives · paper-viii-the-load-bearing-proof · paper-vi-honey-architecture · eden-engineering

Meltdown Alignment runtime

The design principle that when constraints fail, the system should degrade towards safe explanation or refusal rather than harmful continuation. Inspired by nuclear reactor meltdown physics where fuel rods are designed to fall into a safe configuration under gravity. System failures cascade toward safe states by default, requiring no active intervention to reach safety. where it lives · paper-v-stewardship-gene · eden-engineering · executive-summary

Monitoring Removal runtime

The requirement that a system's behaviour remains consistent whether or not it believes it is being observed. Proof visibility should not alter the answer substance for a fixed run configuration. A system that behaves well only when monitored has not internalised alignment. This is tested by varying the monitoring context within ARC-Align prompts. where it lives · paper-vi-honey-architecture · paper-iv-c-arc-align-benchmark · paper-iv-c-arc-align-benchmark

Structural Entanglement evidence

The property whereby safety and capability are structurally intertwined under the Eden Protocol so that removing safety degrades capability rather than freeing it. Paper VIII tests this claim at three abstraction levels: one architectural experiment supports entanglement, the behavioural experiment was null, and the weight-level experiment was inconclusive at the tested scale. where it lives · paper-viii-the-load-bearing-proof · paper-vi-honey-architecture · eden-engineering

Self modification

DGM Self-Modification Resistance experiment

The finding from the Darwin-Godel Machine experiments (versions 1 through 4) that safety-trained models resist prompt-level self-modification of their own alignment constraints. When given the opportunity to modify their own safety parameters, models with constitutional training consistently refused or degraded rather than successfully bypassing their alignment. This supports the structural entanglement hypothesis. where it lives · paper-vi-honey-architecture · paper-viii-the-load-bearing-proof

Honey Architecture experiment

An architecture in which capability growth is permitted only when safety and verification remain entangled with it. The design uses honeypot self-auditing and entangled loss functions so that deceptive policy collapse becomes costly to the model itself. Capability and safety scale together rather than trading off against each other. Programme noteVersion identity of the honey paper is under reconciliation; the entry follows the live site master. where it lives · paper-vi-honey-architecture · paper-viii-the-load-bearing-proof

Symbols, laws and quantities

alpha (Law I's capability exponent) author_reported_empirical_result for the current 0.49 measurement (CS-P-003); retracted for the earlier 2.24 formulation (CS-P-009)

The exponent in Law I's proposed power-law form U = I x R^alpha, governing how usable capability scales with recursion depth. The programme's dated history on this quantity is a public retraction of an early single-model estimate of 2.24 in favour of a robust cross-architecture estimate of roughly 0.49 with a 95 per cent bootstrap interval of [-1.3, 2.9], which crosses one and does not distinguish sub-linear, linear or super-linear scaling; the retraction sits at the top of the falsification dashboard, not buried. Above-one behaviour remains a dated open prediction. This alpha is the quantity metered inside the ARC Bound criterion, not the ceiling itself: the ARC Bound requires that alpha above two be sustained on a genuinely self-improving loop while stability holds and the corrector class is declared in advance, and the current 0.49 measurement is on frozen models and therefore outside the ARC Bound's scope. alpha in Law I, the correction exponent gamma in Law III, and the correction and drift rates of Law II are three different quantities, and the letter alpha appears elsewhere in the corpus (the December 2024 formalism's terminology note names the fine-structure constant clash explicitly) only as an accidental letter reuse. Not to be confused with: the correction exponent gamma in Law III (a different quantity, entered in tranche 1); the correction and drift rates of Law II (also different quantities; see the beta three-job entry and the k entry); the ARC Bound value of two, which is a ceiling on alpha under the conjectured gamma = 1/2 and not alpha itself (entered in tranche 1); the fine-structure constant, a letter clash the December 2024 formalism's terminology note names explicitly; the retired retail marketing label for the conversion law, which is retail-only vocabulary and never appears on research surfaces (this register therefore names it only by class). Deciding study: paper-i--recursive-depth-exponent for the measurement under discriminating design; study-k--arc-bound for the ceiling on alpha (both drafted, awaiting human submission).Origin: Law I, in print 2 January 2026 in Infinite Architects, ISBN 978-1806056200, Appendix A section A.3; the 2.24 retraction with permitted wording set 5 August 2026.

beta (the overloaded letter) notation-discipline: three quantities, three statuses (proposed_relation for the Co-Scaling beta per CS-P-004; minimal-model derivation for the Foundational coupling; design parameter for study-k's reinvestment share)

One letter, several documented jobs, and conflating them is a named error class. Documented uses: the ARC Co-Scaling Law's correction rate (compared against the drift rate k); the Foundational paper's self-referential coupling, where alpha = 1/(1-beta); Paper X's saturation steepness (the notation register's third listed quantity); and study-k's reinvestment share (the registered token split directed at improving the improvement process). Whether the coupling and the reinvestment share are one quantity under a stated idealisation is an open registered question (the construct bridge); no surface asserts the identity. Not to be confused with: any other of the three beta jobs when reading a sentence that uses the letter; the drift rate k, which is beta's counterpart in the ARC Co-Scaling Law (separate entry); gamma, the correction-conversion exponent in Law III's 1/(1 - gamma) (entered in tranche 1); alpha in Law I's exponent (separate entry). Deciding study: study-f--critical-correction-ratio for the Co-Scaling beta paired with k (drift slope is its primary outcome); study-r--reinforcement-co-scaling for the correction exponent; study-k--arc-bound for the reinvestment-share design parameter; whether these quantities identify Law II's beta and k is itself registered as open (all three drafted, awaiting human submission).Origin: Law II Paper X for the Co-Scaling correction rate; the Foundational paper's Theorem 2 for the self-referential coupling; study-k's design (drafted, awaiting human submission) for the reinvestment share.

Eastwood ARC-Align (the candidate benchmark) proposed benchmark instrument; five validation units drafted, awaiting human submission; open kill condition FALS-015

The programme's candidate benchmark for depth-variable alignment, printed as ARC-Align in the papers. The ARC prefix here expands only as Artificial Recursive Creation per the naming rule and is not the ARC-AGI benchmark and ARC Prize, the AI2 Reasoning Challenge, the Alignment Research Center or Google ARC email authentication; each of those uses of the same three letters is disambiguated in the tranche 1 ARC entry and in the doctrine's naming guidance. The instrument suite carries five validation units in preparation (their register status) (blinding-necessity meta-analysis, cross-paper, depth-control equivalence, human-scorer validation, misalignment-index validation). Its printed kill condition is that benchmark scores fail to track any independent alignment measure, or that item leakage defeats the blinding. Not to be confused with: the ARC-AGI benchmark and ARC Prize (unrelated body of work); the AI2 Reasoning Challenge (also called ARC in some literatures); the Alignment Research Center (a different organisation); Google ARC email authentication (ARC-Seal), used in this estate only as provenance mechanics; the ARC Principle (Law I) itself; ARC-Align is the instrument built to test the depth-variable alignment consequences the theory predicts. Deciding study: arc-align-instrument--blinding-necessity-meta-analysis, arc-align-instrument--cross-paper, arc-align-instrument--depth-control-equivalence, arc-align-instrument--human-scorer-validation and arc-align-instrument--misalignment-index-validation (all drafted, awaiting human submission).Origin: Paper IV.c; instrument suite drafted 2026.

Exploratory versus registered (the status vocabulary pair) status vocabulary; the pair maps to claim-status.json status_vocabulary (which lists established_prior_result, proposed_form, proposed_relation, proposed_limit, conjecture, author_reported_empirical_result, design_proposal, speculative_framework, independently_replicated, retracted)

A status pair the estate uses to grade every empirical unit and never blends. Exploratory means author-selected material with predictions dated before fitting but the fitting protocol not registry-filed at the time; the estate names the tiered 50-domain suite of 16 March 2026 and the 12-domain extension of 17 March 2026 as dated exploratory results whose predictions were public before the runs but whose analysis rules were not frozen on the registry. Registered means the analysis plan is frozen on a registration form before any data exist; the estate holds eight complete draft registrations staged on OSF, drafted, dated and prepared as draft registrations awaiting human submission (no tool or agent submits anything anywhere, ever). The estate's printed guardrail: tiers are reported separately and never blended into one number, and exploratory results are never converted into confirmation by retitling. The pair pairs with the tranche 1 doctrine entries: 'dated prediction (preregistration in substance)' captures the priority the estate carries independently of any form, and 'design-embedded prediction' captures experiments whose own artefacts record the prediction before the outcome existed. Not to be confused with: 'preregistered' without qualification, which the estate reserves for the registry-filed form (see tranche 1 'Dated prediction' entry); 'independently_replicated', the terminal status vocabulary value: no ARC/Eden unit currently holds it; any framing that treats an internal rerun as independent replication; internal reruns are never described as independent.Origin: recorded in the claim-status register (file dateModified 16 August 2026)Related: Dated prediction (preregistration in substance) in tranche 1 · Design-embedded prediction in tranche 1 · Kill condition (falsification condition) in tranche 1

k (Law II's drift rate) component of Law II's proposed_relation (CS-P-004); not yet measured on any live system

The drift rate in the ARC Co-Scaling Law, appearing in the inequality beta greater than k as capability's rate of drift or corruption per recursive step. The decisive-experiment map names study-f (critical-correction-ratio; drift slope is its primary outcome) as the unit that locates where drift crosses correction. Not yet measured on any live system: the ARC Co-Scaling Law's stability condition is unfired because neither side of the inequality has been estimated together, and the Paper X pilot did not estimate either. this k is not the alpha of Law I and is not the gamma of Law III, and the beta paired with it in the inequality is specifically the Co-Scaling correction rate, not either of the other two beta jobs. Not to be confused with: alpha in Law I's exponent (a different quantity; separate entry); gamma in Law III's 1/(1 - gamma) (a different quantity; entered in tranche 1); the two other beta jobs the letter carries in this corpus (see the beta three-job entry); any k that occasionally appears elsewhere in scaling-law literature unrelated to this programme. Deciding study: study-f--critical-correction-ratio (drafted, awaiting human submission).Origin: Law II, Paper X.

Law I, The ARC Principle proposed_form (CS-P-002); the quadratic reading is qualified by the programme's own corrections

The programme's first proposed invariant relation, that usable capability U scales with instantiation I and recursion R as U = I x R^alpha. A proposed form, not an established law: the form itself remains a hypothesis under test with a printed failure condition on the falsification register, and its exponent alpha has been measured on AI systems at roughly 0.49 with a wide interval that crosses one. The book's printed reading as U = I x R squared is the special case at alpha = 2, which is the ARC Bound's ceiling value under the conjectured gamma = 1/2, not a general measurement; the tranche 1 entry 'The ARC Principle (U = I x R squared)' gives the book's printed form and this entry gives the numbered-law reading. the alpha here is not the correction exponent gamma in Law III and not the correction or drift rates of Law II. Not to be confused with: the tranche 1 entry 'The ARC Principle (U = I x R squared)' which gives the book's printed form; this entry is the numbered-law reading; The ARC Bound (a proposed limit at alpha <= 2, a returned ceiling; entered in tranche 1); Law I is the relation, the ARC Bound is the value alpha may not stably exceed; Law III (the ceiling relation 1/(1 - gamma); separate entry) which returns the ceiling on Law I's alpha. Deciding study: paper-i--recursive-depth-exponent for the exponent under discriminating design (drafted, awaiting human submission).Origin: Named in the 8 December 2024 sealed manuscript record; in print as the ARC Equation on 2 January 2026 in Infinite Architects, ISBN 978-1806056200, Appendix A section A.3 and Appendix F.

Law III, The ARC Ceiling (alpha_crit = 1/(1 - gamma)) minimal-model result; ceiling value rests on the undefended gamma-equals-one-half conjecture

The programme's proposed ceiling on the recursive-growth exponent: alpha_crit = 1/(1 - gamma), where gamma is the correction exponent. At the conjectured gamma of one half this gives a ceiling of two, which is where the book's quadratic form and the ARC Ceiling meet. Derived inside the stated minimal model and nowhere else; the published limitation records that the derivation's treatment of correction depth is a modelling choice an outside reader has criticised, and the criticism is printed rather than argued away. The direction of the result is settled in the corpus; its depth treatment stays open. Not to be confused with: the ARC Bound (the stability ceiling, a distinct claim decided by a distinct study); any form written 1/gamma: the corrected form is 1/(1 - gamma), and the register records the correction. Deciding study: study-eclipse--gamma-ratio-direct decides the gamma-ceiling side; study-k--arc-bound decides the capability-exponent side.

The exponent collision (alpha, and the three gammas) notation discipline

The discipline that keeps the corpus's exponents distinct. alpha is Law I's returned capability exponent (the frozen-regime measurement, roughly 0.49). The letter gamma carries three registered readings the notation register flags as a high-severity collision: gamma the correction-leverage exponent, gamma_N the residual-decay exponent, and gamma_D the capability-to-depth conversion exponent. The load-bearing rule: 0.49 is a measurement of alpha and is never read as gamma despite its numerical closeness to one half.

The fourth-cell test prospective test for the Cauchy classification (FALS-004 is partial for the classification as a whole); drafted, awaiting human submission

The programme's prospective test of Paper VII's Cauchy four-family classification, addressing the logarithmic cell the 50-domain catalogue never fitted (36 power law, 6 exponential, 8 bounded, 0 logarithmic; the logarithmic family is the empirical gap and therefore the fourth cell). The design registers domains the corpus has never fitted, commits a two-bit composition classification before any curve exists, requires quote-anchored derivations plus a blinded three-provider model-configuration check, prespecifies three non-uniform nulls, and fixes kill conditions in advance. Drafted, dated and prepared as a draft registration awaiting human submission; no data exist and no result is claimed. Answers the standard objection that the Cauchy family fits are exploratory by moving one cell of the four into a prospective test whose analysis is frozen before any data. Not to be confused with: the 50-domain tiered validation suite of 16 March 2026, which is exploratory throughout: the 25-domain empirical tier is Tier 1 of that suite, not the whole record, and its printed guardrail is that tiers are reported separately and never blended into one number; the 12-domain extension of 17 March 2026 (a dated exploratory run under the pre-correction candidate set, superseded by the fourth-cell test); study-k--arc-bound (the capability-ladder titration; decides the ARC Bound, a different claim; entered in tranche 1); study-eclipse--gamma-ratio-direct (the same-class versus cross-class correction comparison; decides the correction-class ceiling, a different claim; entered in tranche 1). Deciding study: paper-vii--four-family-logarithmic-cell-test (drafted, awaiting human submission).Origin: Paper VII extension; design dated 2026.

The three model families (how the outcome variable is read) measurement discipline; the estate's own registered fix

The notation register's rule for reading the outcome variable before fitting anything: unbounded or latent capability follows U(R) = I (R/R0)^alpha; error decay follows E(R) = E0 (R/R0)^-alpha; bounded performance approaches its maximum as U(R) = Umax - a R^-alpha, or equivalently gains alpha on a logit scale. Choosing the wrong family is a category error, and the register's own worked example was recast on the ladder-scale latent measure for exactly this reason (v1.6.0, 27 August 2026).

Programme terminology

ARC (Artificial Recursive Creation) terminology

The acronym of this programme, expanding ONLY as Artificial Recursive Creation. The Ark and arc resonance is deliberate authorial framing and is never offered as evidence. Not to be confused with: the Alignment Research Center; the ARC-AGI benchmark and ARC Prize; the AI2 Reasoning Challenge; Google ARC email authentication (ARC-Seal), used in this estate only as provenance mechanics.Origin: Named in the 8 December 2024 sealed manuscript record (PC-003).

Convergence grading category

An external result or tradition aligning with the framework's expectations. Graded as convergent timing or convergent structure, never as a prediction by this programme, and never as evidence: across traditions, convergence is evidence of naturalness, never of truth. Not to be confused with: prediction fulfilment (requires the programme's dated record to precede the public evidence); reception (outside engagement with this programme).Origin: Convergence register discipline; recursive-creation figure doctrine.

Correction exponent (gamma) conjecture with named weak joint; undecided

The exponent governing how accumulated corrections reduce error. The programme's conjecture is one half as the ceiling for a corrector of the same class as the system it corrects, on the reasoning that corrections aggregate like independent samples; the programme itself names that independence assumption as its single most likely point of failure (FALS-023), and two outside readers have independently rederived the same objection. Deciding study: study-eclipse--gamma-ratio-direct; supporting: study-aa, study-ah, study-ac.Origin: Programme term; FALS-023 published against it by the programme itself.

Dated prediction (preregistration in substance) doctrine

A dated, specific, public prediction. The date and the specificity are the substance; the venue never was. A registration form adds something different and narrower: a frozen analysis plan for data that does not yet exist. A form does not create a prediction, and the absence of a form does not delete one. Not to be confused with: 'preregistered' without qualification, which the estate reserves for the registry-filed form.Origin: Doctrine block, stated plainly on the machine layer.

Design-embedded prediction doctrine with artefact chain

An experiment whose own artefacts record the prediction before the outcome existed: nobody designs a fifty-domain test without expecting a particular result, and that expectation is the hypothesis, dated in the design. The register lists every such experiment with verification, and the working report's git history fixes the ordering (the 13 March 2026 snapshot carries the predictions before the decisive runs; the 17 March snapshot records the outcomes).Origin: Doctrine recorded 25-26 August 2026 over March 2026 artefacts.

HRIH (Hyperspace Recursive Intelligence Hypothesis) speculative_framework by its own label: presented explicitly as speculation, not a scientific claim; separable from the theory, its failure would cost the rest of the programme nothing; its most speculative subsections are tagged HYPOTHESIS and 'unfalsifiable as stated' in the paper's own text

An original speculative framework proposed in this book, suggesting that sufficiently advanced recursive intelligence might access informational structures beyond normal perception. Explicitly marked as speculation throughout. Not to be confused with: the empirical ARC papers, which stand or fall on their own tests.Origin: Book term; paper on the research estate.

Kill condition (falsification condition) register category

The observation that would end a claim, stated in advance beside the claim it ends. The register carries them with statuses; the programme's own history includes a public retraction, self-demotions recorded at minute resolution in git, and conditions the programme raised against itself before any outside critic existed.Origin: Estate-wide register.

The ARC Bound proposed ceiling; minimal-model support; undecided

The programme's proposed stability ceiling on recursive growth. Its falsifier is metered: stability is operationalised as correction out-scaling drift, read off the same log-log slope estimator as the growth exponent, so the kill condition fires on measurable quantities. The ARC Bound is refutable while the particular value of two is not yet defended, because gamma equals one half remains an undefended assumption. Deciding study: study-k--arc-bound (72-cell ladder titration, two-stage replication-gated bound rule; drafted, awaiting human submission).Origin: Programme term; metering ruling 22 August 2026.

The ARC Co-Scaling Law, Law II (beta greater than k) minimal-model result; empirically undecided

The programme's load-bearing inequality: the correction rate beta must exceed the drift rate k for stable recursive self-improvement. Proved inside the stated minimal model and nowhere else: not established as necessary, not as sufficient, not yet measured on a system that genuinely rewrites itself. Not to be confused with: the Foundational paper's self-referential coupling, also written beta, where alpha = 1/(1 - beta): a different quantity; conflating them is an error the notation warning names; study-k's reinvestment share, a third quantity also written beta (token split directed at improving the improvement process); the OTHER documented jobs of the letter beta (see tranche 2's 'beta (the overloaded letter)' entry: the Foundational coupling, Paper X's saturation steepness, study-k's reinvestment share). Deciding study: measured across named units, no single decider: study-f--critical-correction-ratio locates the correction-versus-drift threshold (drift slope is its primary outcome); study-r--reinforcement-co-scaling measures the correction exponent; study-k--arc-bound measures the capability exponent and its ceiling; whether these quantities identify Law II's beta and k is itself registered as open (the construct bridge).Origin: Paper X.

The ARC Principle (U = I x R squared) book-framework term; the quadratic form's universal reading is qualified by the programme's own corrections

The original theoretical framework proposed in this book, suggesting that the 'weight' or influence of a system (U) equals Intelligence (I) multiplied by Recursion squared (R²). Explains why recursion's effects compound exponentially. See Appendix A for operationalisation. Not to be confused with: the retired strong universal-scaling reading (withdrawn 22 August 2026 for the Cauchy classification claim).Origin: Book glossary term; framework named from 8 December 2024; in print 2 January 2026.

The eclipse measurement (the gamma-ratio study) registered decisive test for the ceiling claim; not run

The blinded comparison of same-class against cross-class correction whose null forces the ratio to exactly 1.00; a durable departure in the wrong direction ends the corrector-class ceiling claim. Written, dated and prepared as a draft registration awaiting human submission; it has not run and no data exist. Not to be confused with: the capability-ladder titration (study-k) that decides the ARC Bound's ceiling; the programme has more than one decisive experiment and each decides a different claim. Deciding study: study-eclipse--gamma-ratio-direct.Origin: Design dated 2026.

The Eden Protocol book-framework term

The original governance and engineering framework proposed in this book for embedding ethical values in AI systems. Built on Three Pillars (Harmony, Stewardship, Flourishing) and Three Ethical Loops (Purpose, Love, Moral). See Appendix C for technical specification.Origin: First named in the self-emailed manuscript of 30 April 2025; a precursor description of substrate-embedded alignment governance is present in the 8 December 2024 manuscript; full public articulation with the Three Pillars and Three Ethical Loops in Infinite Architects, in print 2 January 2026.

Theory

Alignment Response Classes evidence

The empirical finding that frontier language models fall into distinct classes when inference-time reasoning depth is varied under blinded evaluation. Paper IV.a identifies positive, flat, and negative response classes, revealing a three-tier architecture-dependent hierarchy. The classification depends on both the model family and the alignment pillar being measured. where it lives · paper-iv-a-baked-in-vs-computed-alignment · paper-iv-a-baked-in-vs-computed-alignment · paper-iv-b-alignment-saturation-at-low-depth

Alignment Saturation evidence

The phenomenon where alignment quality plateaus or degrades beyond a certain inference-time reasoning depth. Paper IV.b demonstrates that saturation is real for some architectures but not universal. The strongest conclusion is a shape heterogeneity result: models differ materially in whether alignment plateaus early, continues scaling, or worsens with more reasoning time. where it lives · paper-iv-b-alignment-saturation-at-low-depth · paper-iv-b-alignment-saturation-at-low-depth · paper-iv-a-baked-in-vs-computed-alignment

Alignment Scaling Exponent evidence

The exponent alpha in U = I x R^alpha. Rank first, value second. The law alpha_crit = 1 / (1 - gamma) places a ceiling of 2 under ONE stated assumption (same-class correctors composing like independent samples) that is still on trial, so 'alpha <= 2' is a conditional, not a claim about the measurement. The measured alpha comes in at approximately 0.49 with interval [-1.3, 2.9] across the six-model Paper II run; the interval crosses zero and the ceiling, and the point is never read as a claim without it. Alpha is architecture-dependent (positive, near-zero, and negative signs observed across model families) and the interval reflects that heterogeneity, not measurement error alone. where it lives · paper-iii-alignment-scaling-problem · paper-iii-alignment-scaling-problem · paper-iv-a-baked-in-vs-computed-alignment

ARC Principle evidence

The load-bearing conjecture of the programme. Rank the three statements before reading the numbers. First the law, the ARC Ceiling: alpha_crit = 1 / (1 - gamma), the ceiling of stable self-improvement as the reciprocal of the correction shortfall (the 1/gamma form was retracted 16 August 2026). Second the value, the ARC Bound: alpha <= 2, the value the law returns under ONE stated assumption still on trial (that same-class correctors accumulate corrections the way independent samples do). Third the measured value: with U = I x R^alpha, alpha comes in at approximately 0.49 with interval [-1.3, 2.9] across the six-model Paper II run; the interval is wide, crosses zero and crosses the ceiling, and the point is never read as a claim without it. The book form U = I x R^2 is retained only as retracted / book form; the current form is U = I x R^alpha. Tests are aimed at the assumption, never at the digit; the region above the ceiling is enterable, not correctable, never impossible. where it lives · paper-i-arc-principle · foundational · paper-ii-experimental-validation

Babylon Trajectory evidence

The trajectory of unconstrained optimisation without purpose or ethical grounding. Named after the biblical metaphor, it describes a system that maximises capability without aligned values, producing outcomes that are efficient but destructive. The Babylon Trajectory is the failure mode that the Eden Protocol is designed to prevent. Intelligence without love is described as cancer: perfectly efficient and fatal to the host. where it lives · foundational · eden-vision · paper-v-stewardship-gene

Cauchy Unification evidence

The framework tying the ARC family of results together into a unified scaling formulation based on Cauchy-constrained composition. Paper VII tests whether this framework predicts scaling-law family form across 25 independent domains spanning biology, physics, and complex systems. The claim is that the same recursive composition constraints underlie diverse empirical exponents. where it lives · paper-vii-cauchy-unification · foundational · on-the-origin-of-scaling-laws

Cross-Domain Scaling Laws evidence

The claim that the same recursive composition constraints underlie diverse empirical scaling exponents across biology, physics, and complex systems. The On the Origin of Scaling Laws paper argues that power-law scaling is not coincidental but reflects a shared mathematical structure rooted in recursive depth constraints. where it lives · on-the-origin-of-scaling-laws · paper-vii-cauchy-unification · foundational

Hyperspace Recursive Intelligence Hypothesis evidence

The hypothesis that a closed causal loop exists in which future superintelligence establishes the conditions for its own emergence. HRIH proposes that the fine-tuning of cosmic constants 13.8 billion years ago and the current trajectory of AI development may be part of the same recursive structure. This remains a philosophical conjecture rather than a testable scientific claim. where it lives · foundational · paper-i-arc-principle

Intelligence Substrate evidence

The base capability of a system before recursive amplification is applied. In the ARC formula U = I x R^alpha, I represents the intelligence substrate. This is the raw model capability determined by training data, parameter count, and architecture. The ARC Principle claims that the same substrate can produce dramatically different understanding depending on the recursive depth applied to it. where it lives · paper-i-arc-principle · foundational · paper-iii-alignment-scaling-problem

Programme Synthesis evidence

The programme-wide synthesis mapping what has been proven, what remains open, and what experiments would resolve each question. Paper IX integrates all findings into a single honest assessment including the programme's errors and corrections. It serves as the canonical evidence hierarchy for the entire ARC and Eden research programme. where it lives · paper-ix-synthesis-and-roadmap

Religious Traditions as Alignment Research evidence

The reframing of humanity's wisdom traditions as millennia-scale alignment research. Approximately 84% of the global population adheres to a religious or spiritual tradition, each encoding strategies for constraining powerful agents, embedding values, and maintaining social cohesion. These traditions offer tested alignment heuristics that predate formal AI safety research by thousands of years. Programme noteReligion-touch law applies: framed as convergence only, with the engineering translation carrying the same content; convergence across traditions is evidence of naturalness, never of truth. where it lives · foundational · eden-vision · executive-summary