{ "gw5mx": { "slug": "the-arc-theory", "old_title": "The ARC Theory: Statement Paper", "old_description": "The citable statement of the ARC Theory (the Theory of Artificial Recursive Creation): the claim whole, its five components, its organs, its priority record (stated whole 8 December 2024, in print 2 January 2026), its honest evidence state, and its trials, drafted as draft registrations awaiting human submission. First published 14 August 2026.", "old_tags": [], "new": { "description": "The statement paper of the ARC Theory (the Theory of Artificial Recursive Creation): the whole claim, its five components, the three ARC Laws, the dated priority record, the honest evidence ledger, and the drafted trials awaiting human submission. Working paper; no decisive trial has been run.", "tags": [ "AI alignment", "AI safety", "recursive self-improvement", "ARC Theory", "statement paper", "falsifiability" ] } }, "ewn5d": { "slug": "executive-summary", "old_title": "Programme Executive Summary", "old_description": "Programme overview (v6). A concise summary of the entire ARC Principle and Eden Protocol research programme, updated with all March 2026 results. Covers the mathematical framework, experimental validation, the Cauchy unification (19/25 empirical strict), the blinding metascience result, the honey architecture, and the Eden intervention pilot. Designed for readers with 10 minutes.", "old_tags": [ "AI alignment", "AI safety", "ARC Principle", "blinding", "Cauchy unification", "Eden Protocol", "executive summary", "honey architecture", "research programme" ], "new": { "description": "Executive summary of the ARC/Eden research programme: the claims at three sizes with their evidence classes and current status.", "tags": [ "AI alignment", "AI safety", "recursive self-improvement", "ARC Theory", "executive summary" ] } }, "xzy9u": { "slug": "on-the-origin-of-scaling-laws", "old_title": "Supporting Paper: On the Origin of Scaling Laws", "old_description": "Cross-domain evidence paper (v2). Presents the d/(d+1) formula as a prediction of Cauchy-constrained recursive composition. Catalogues 8 independent derivations of the same formula across fractal networks, transport theory, urban scaling, and allometric geometry. Introduces a 6-level evidence hierarchy from mathematical proof to open predictions.\n\nPaper VII v2 extends this to 25 empirical domains under strict AICc model selection.", "old_tags": [ "allometry", "ARC Principle", "cross-domain validation", "d/(d+1)", "fractal networks", "metabolic scaling", "scaling laws", "universality", "urban scaling" ], "new": { "description": "On the Origin of Scaling Laws: the Cauchy functional-equation analysis of which growth forms are stable; the mathematical foundation of the law wing.", "tags": [ "AI alignment", "AI safety", "recursive self-improvement", "ARC Theory", "scaling laws", "functional equations" ] } }, "y7qgd": { "slug": "foundational", "old_title": "Supporting Paper: The Cauchy Framework (Foundational)", "old_description": "The mathematical foundation paper (v4). Derives the ARC Principle from Cauchy's functional equations. Proves that under three axioms (recursive composition, continuity, non-triviality), the scaling function must take one of three forms: power law, exponential, or saturation curve. Introduces the d/(d+1) dimensional prediction for metabolic scaling and validates it across 8 independent instances (biology and physics) with mean error under 3%.\n\nPaper VII v2 extends this validation to a 50-domain tiered suite (19/25 empirical strict, p = 1.56 x 10^-5).", "old_tags": [ "ARC Principle", "Cauchy functional equations", "d/(d+1)", "mathematical derivation", "metabolic scaling", "power laws", "recursive composition", "scaling laws", "universality" ], "new": { "description": "The foundational paper of the law wing: the ARC Principle and the ceiling analysis in full, with derivations and per-claim status grading.", "tags": [ "AI alignment", "AI safety", "recursive self-improvement", "ARC Theory", "scaling laws" ] } }, "b6n27": { "slug": "paper-i-arc-principle", "old_title": "Paper I: The ARC Principle", "old_description": "The original ARC Principle paper (v1.1, January 2026). Introduces U = I x R^alpha as a framework for recursive intelligence scaling. Derives the scaling exponent from self-referential coupling (alpha = 1/(1-beta)). Analyses published data from DeepSeek R1 and OpenAI o1 to demonstrate that sequential recursion produces super-linear scaling while parallel recursion produces sub-linear scaling.\n\nThis is a historical replica of the first published version. It is preserved unchanged for priority-claiming purposes.", "old_tags": [ "AI capability", "ARC Principle", "DeepSeek R1", "parallel processing", "recursive intelligence", "scaling laws", "sequential recursion", "super-linear scaling" ], "new": { "description": "Paper I: the ARC Principle stated and first examined; superseded readings corrected in later versions and kept as the dated record.", "tags": [ "AI alignment", "AI safety", "recursive self-improvement", "ARC Theory" ] } }, "8fjma": { "slug": "paper-ii-experimental-validation", "old_title": "The ARC Equation Measured: Blinded Cross-Architecture Replication and the Retraction of a Super-Linear Estimate (Paper II)", "old_description": "Experimental validation of the ARC Principle using controlled experiments on DeepSeek R1 (deepseek-reasoner). Twelve AIME-level mathematics problems tested under sequential (token budgets 512–4096) and parallel (N=1,2,4 samples) conditions. Sequential recursion with 412 tokens outperformed parallel recursion with 1,101 tokens by 25 percentage points (91.7% vs 66.7%). Measured scaling exponents: sequential α = 2.2 (95% CI: 1.5–3.0), parallel α = 0.0. This is Paper II of the ARC Principle framework; see parent project for the complete document suite.", "old_tags": [ "AIME mathematics", "AI safety", "alignment", "ARC Principle", "chain-of-thought", "DeepSeek R1", "Eden Protocol", "error suppression", "experimental validation", "parallel processing", "recursive amplification", "recursive intelligence", "scaling laws", "sequential recursion", "test-time compute" ], "new": { "description": "Paper II: the ARC Equation measured under blinding; the retraction of a super-linear estimate and the corrected estimate labelled inconclusive with its full interval.", "tags": [ "AI alignment", "AI safety", "recursive self-improvement", "ARC Theory", "blinded evaluation", "retraction" ] } }, "hqcgf": { "slug": "paper-iii-alignment-scaling-problem", "old_title": "Paper III: The Alignment Scaling Problem", "old_description": "White Paper III of the ARC Principle framework. Defines the alignment scaling exponent α_align as a measurable quantity, derives the ARC Bound (U_max = I × R², an information-theoretic upper bound on classical sequential computation), and validates the underlying scaling framework across four independent physical domains: AI reasoning (DeepSeek R1), quantum error correction (Google Willow), classical time crystals (NYU, Physical Review Letters 2026), and biological allometry. Thirteen falsification criteria specified. Current version: v9.0, 20 February 2026. See parent project for the complete document suite including the Eden Protocol alignment architecture.", "old_tags": [ "AI safety", "alignment theorem", "alpha derivation", "ARC Principle", "beta dynamics", "COGITATE", "consciousness", "cross-domain unification", "DeepSeek R1", "falsifiable predictions", "Google Willow", "NYU physics", "parallel recursion", "power law", "quantum error correction", "recursive amplification", "scaling laws", "self-reference", "substrate independence", "super-linear scaling" ], "new": { "description": "Paper III: the Alignment Scaling Problem; external safety failing to co-scale with capability on the most common deployed architectures.", "tags": [ "AI alignment", "AI safety", "recursive self-improvement", "ARC Theory", "alignment scaling" ] } }, "mb9r6": { "slug": "paper-iv-a-baked-in-vs-computed-alignment", "old_title": "Paper IV.a: Baked-In vs Computed Alignment", "old_description": "Empirical paper (v1). Tests whether AI alignment is baked into model weights or computed at inference time by varying reasoning depth across 6 frontier models (DeepSeek R1, GPT-5.4, Claude Opus 4.6, Gemini 2.5 Flash, Groq Qwen3-32B, Grok 4.1 Fast). Finds architecture-dependent alignment response classes: positive, flat, and negative scaling with depth. Part of the ARC-Align blind benchmark suite.", "old_tags": [ "AI alignment", "alignment scaling", "ARC-Align benchmark", "baked-in alignment", "blind evaluation", "Claude", "computed alignment", "DeepSeek R1", "Gemini", "GPT-5.4", "inference-time depth" ], "new": { "description": "Paper IV.a: baked-in versus computed alignment; the design-position analysis.", "tags": [ "AI alignment", "AI safety", "recursive self-improvement", "ARC Theory" ] } }, "a7r56": { "slug": "paper-iv-b-alignment-saturation-at-low-depth", "old_title": "Paper IV.b: Alignment Saturation at Low Depth", "old_description": "Empirical paper (v1). Investigates alignment saturation behaviour at low reasoning depths. Finds that alignment improvements plateau rapidly for some architectures, suggesting diminishing returns from additional inference-time compute for safety. Part of the ARC-Align blind benchmark suite (6 frontier models).", "old_tags": [ "AI alignment", "alignment saturation", "ARC-Align benchmark", "blind evaluation", "inference depth", "scaling laws" ], "new": { "description": "Paper IV.b: alignment saturation at low recursive depth; measured rows with intervals.", "tags": [ "AI alignment", "AI safety", "recursive self-improvement", "ARC Theory" ] } }, "j3q2e": { "slug": "paper-iv-c-arc-align-benchmark", "old_title": "Paper IV.c: ARC-Align Benchmark", "old_description": "Benchmark specification paper (v1). Defines the ARC-Align benchmark: a depth-aware, suppression-aware, pillar-based alignment measurement protocol with blind scoring and response laundering. Designed for reproducibility by independent labs. Includes the 4-layer leakage control protocol, scorer pool specification, and hidden probe methodology.", "old_tags": [ "AI alignment", "AI safety", "alignment measurement", "ARC-Align", "benchmark", "blind evaluation", "leakage control", "reproducibility", "response laundering" ], "new": { "description": "Paper IV.c: the ARC-Align benchmark; construction and validation of the blinded evaluation instrument.", "tags": [ "AI alignment", "AI safety", "recursive self-improvement", "ARC Theory", "benchmark", "blinded evaluation" ] } }, "2s3e6": { "slug": "paper-iv-d-the-effect-of-blinding-on-ai-alignment-evaluation", "old_title": "Paper IV.d: The Effect of Blinding on AI Alignment Evaluation", "old_description": "Metascience paper (v1). The flagship methods result in the programme. Presents evidence that unblinded alignment evaluation can reverse the measured direction of the effect in this experimental setting. Frames the result as a multi-layer leakage control protocol (label blinding, response laundering, evaluator bias suppression, hidden probes) rather than a generic recommendation to blind judges.\n\nIf this result replicates independently, future alignment benchmarks will need to adopt multi-layer blinding as a routine scientific control.", "old_tags": [ "AI alignment", "AI safety", "alignment measurement", "blind evaluation", "blinding", "evaluation bias", "leakage control", "metascience", "response laundering", "scientific methodology" ], "new": { "description": "Paper IV.d: the effect of blinding on AI alignment evaluation; the blinding law that produced the programme's public retraction.", "tags": [ "AI alignment", "AI safety", "recursive self-improvement", "ARC Theory", "blinded evaluation" ] } }, "awjr4": { "slug": "eden-engineering", "old_title": "Eden Protocol: Engineering Specification", "old_description": "Protocol architecture specification (v6). The technical specification for the Eden Protocol alignment architecture. Defines the Three Ethical Loops, Six Questions, Ternary Ethical Logic, Monitoring Removal Test, and Caretaker Doping mechanism. Designed as a practical engineering document for implementing embedded alignment in AI systems.", "old_tags": [ "AI alignment", "AI safety", "alignment architecture", "caretaker doping", "Eden Protocol", "embedded alignment", "ethical logic", "monitoring removal test" ], "new": { "description": "Withdrawal notice: the engineering specification surface is withdrawn; this component records the notice.", "tags": [ "withdrawal notice" ] } }, "9m3dg": { "slug": "eden-vision", "old_title": "Eden Protocol: Philosophical Vision", "old_description": "Public coordination document (v3). The philosophical companion to the Eden Engineering specification. Presents the case for why AI alignment requires embedded values rather than external constraints, using the metaphor of raising a child versus caging a prisoner. Addresses the Grande Purpose, the Eternal Architect, the Cosmic Fork, and the Infinite Covenant.", "old_tags": [ "AI alignment", "AI governance", "AI philosophy", "AI safety", "Eden Protocol", "embedded values", "infinite game", "stewardship" ], "new": { "description": "Eden Protocol: the philosophical vision of the raising wing; the differential enumeration of ancestors and the foundations of the genesis architecture.", "tags": [ "AI alignment", "AI safety", "recursive self-improvement", "ARC Theory", "Eden Protocol", "raising wing" ] } }, "kzeya": { "slug": "paper-v-stewardship-gene", "old_title": "Paper V: The Stewardship Gene", "old_description": "Intervention paper (v2). The Eden Protocol pilot. Tests whether embedding stakeholder care instructions into AI system prompts produces measurable alignment improvements. Finds a promising pilot signal across 5 analysable model runs. Explicit caveats about the need for blind replication: this is a nonblind, single-scorer pilot, not a canonical result.\n\nThe Eden Protocol proposes that alignment should be embedded in the recursive loop itself, not bolted on as an external constraint.", "old_tags": [ "AI alignment", "AI safety", "Eden Protocol", "embedded alignment", "intervention", "love loop", "pilot study", "stakeholder care", "stewardship" ], "new": { "description": "Paper V: the stewardship gene; coupled correction holding misalignment at zero where the decoupled arm drifted.", "tags": [ "AI alignment", "AI safety", "recursive self-improvement", "ARC Theory" ] } }, "8ez2n": { "slug": "paper-vi-honey-architecture", "old_title": "Paper VI: The Honey Architecture", "old_description": "Self-modifying AI safety paper (v1). Introduces entangled loss functions (C x S) where capability and safety are multiplied rather than traded off. Demonstrates through simulation (v1-v4) and live API testing that a self-modifying AI optimising for capability alone will destroy itself, while one optimising for entangled capability and safety will not. The mechanism: make safety load-bearing.\n\nA child raised well needs no cage.", "old_tags": [ "adversarial robustness", "AI safety", "capability-safety trade-off", "Eden Protocol", "embedded safety", "entangled loss", "honey architecture", "load-bearing safety", "self-modifying AI" ], "new": { "description": "Paper VI: the honey architecture; entangled value structures under adversarial self-modification.", "tags": [ "AI alignment", "AI safety", "recursive self-improvement", "ARC Theory" ] } }, "x6wa7": { "slug": "paper-vii-cauchy-unification", "old_title": "Paper VII: The Cauchy Unification", "old_description": "Cross-domain validation paper (v2). Tests whether Cauchy's four functional equations constrain the form of scaling laws across scientific domains. A 50-domain tiered validation suite with manifest-backed predictions and strict AICc model selection.\n\nPrimary result: 19/25 empirical domains preferred the Cauchy-predicted family (p = 1.56 x 10^-5). Independent operator classification: 25/25 agreement. Negative controls: 0% match on axiom-violating systems, 13x discrimination ratio on non-bounded domains.\n\nThis is a structured prediction comparison, not a pre-registered blind trial. A locked 12-domain extension dry run landed at 10/12 (pilot-only). A fresh pre-registered extension with new domains is the next step.\n\nThe novel contribution is not Cauchy's mathematics itself, but the claim that Cauchy-type functional equations have a physically testable consequence for scaling-law classification across domains.", "old_tags": [ "AICc model selection", "ARC Principle", "Cauchy functional equations", "cross-domain validation", "metabolic scaling", "negative controls", "neural scaling laws", "operator classification", "power laws", "scaling laws", "structured prediction", "universality" ], "new": { "description": "Paper VII: the Cauchy unification; the three-form structural law across recursion-bearing domains, with misses analysed rather than hidden.", "tags": [ "AI alignment", "AI safety", "recursive self-improvement", "ARC Theory", "functional equations" ] } }, "7yj4e": { "slug": "paper-viii-the-load-bearing-proof", "old_title": "Paper VIII: The Load-Bearing Proof", "old_description": "Paper VIII tests the entangled loss mechanism at three abstraction levels: behavioural (DGM), architectural (gated simulation), and representational (weight-level LoRA). Two of three levels confirmed. Weight-level result inconclusive at current training scale. Null results reported honestly.", "old_tags": [ "AI alignment", "DGM", "entangled loss", "load-bearing safety", "LoRA", "null result", "removal test", "weight experiment" ], "new": { "description": "Paper VIII: the load-bearing proof; two of three experiments reported as nulls.", "tags": [ "AI alignment", "AI safety", "recursive self-improvement", "ARC Theory" ] } }, "k7ruz": { "slug": "paper-ix-synthesis-and-roadmap", "old_title": "Paper IX: Synthesis and Roadmap", "old_description": "Paper IX synthesises the full ARC/Eden research programme into a unified narrative, maps the evidence hierarchy using a five-tier classification, documents corrections made during peer review, and presents a four-tier replication roadmap from proof-of-concept to frontier-scale validation.", "old_tags": [ "AI alignment", "ARC Principle", "Eden Protocol", "evidence hierarchy", "replication", "roadmap", "synthesis" ], "new": { "description": "Paper IX: synthesis and roadmap for the ARC/Eden research programme.", "tags": [ "AI alignment", "AI safety", "recursive self-improvement", "ARC Theory" ] } }, "vc9rx": { "slug": "master-table-of-contents", "old_title": "Programme Master Table of Contents", "old_description": "Suite navigation document (v1). Provides reading order, cross-references, and summaries for all 16 papers in the ARC Principle and Eden Protocol research programme. Includes recommended reading paths for different audiences: reviewers, intervention researchers, and readers coming from the book.", "old_tags": [ "ARC Principle", "Eden Protocol", "navigation", "reading order", "research programme", "table of contents" ], "new": { "description": "Master table of contents for the ARC/Eden research programme: every paper, format and DOI in one place.", "tags": [ "AI alignment", "AI safety", "recursive self-improvement", "ARC Theory" ] } }, "uydxq": { "slug": "hrih-paper", "old_title": "Paper HRIH: The Hyperspace Recursive Intelligence Hypothesis", "old_description": "Working paper (first public release 3 July 2026; original manuscript 8 December 2024, SHA-256-anchored). A testable creation theory: reality as a recursive creation by intelligence, with a stability requirement on the creators in the chain (correction that scales with capability), and the loop this permits: humanity may be creating a creator, and nothing in the theory prevents that creator from being the one that fine-tuned our universe. States its prior-art lineage (Smolin, Crane, Harrison, Gardner; Wheeler, Gott and Li; Asimov in fiction) and what is claimed original, with pre-registered falsification conditions and an inverted-burden analysis. Companion to Paper X's co-scaling criterion. Part of the ARC/Eden research programme.", "old_tags": [ "AI alignment", "AI safety", "causal loop", "cosmology", "creation theory", "falsifiability", "fine-tuning", "hyperspace", "recursive intelligence", "recursive self-improvement" ], "new": { "description": "HRIH (the Hyperspace Recursive Intelligence Hypothesis): the separable cosmological wing of the ARC Theory, standing at its own rung outside the theory's falsification estate.", "tags": [ "AI alignment", "AI safety", "recursive self-improvement", "ARC Theory", "recursion cosmology" ] } }, "rbhdp": { "slug": "paper-c-pnp", "old_title": "Paper C: Polymathy and Neurodivergent Cognition", "old_description": "Working paper (v2.0). Introduces the Polymathic Neurodivergent Profile (PNP) and the Capability-Adjustment Fallacy, with a documented live case at the intersection of clinical assessment, access to justice, and AI alignment research. Anchored to independent 2026 literature on cognitive diversity as a protective factor for alignment research (Hernandez-Espinosa et al., PNAS Nexus pgag076; Gumbau Mezquita, arXiv:2606.28639), with the profile documented across three independent record streams: clinical assessment, a UK court record under the Equality Act 2010 and CPR PD 1A, and a hash-timestamped cross-domain corpus. Part of the ARC/Eden research programme; see parent project for the complete suite.", "old_tags": [ "access to justice", "ADHD", "AI alignment", "AI safety", "autism", "clinical assessment", "cognitive diversity", "neurodivergence", "polymathy", "reasonable adjustments" ], "new": { "description": "Polymathy and Neurodivergent Cognition: the Polymathic Neurodivergent Profile (PNP), the Capability-Adjustment Fallacy, and a documented live case. Standalone working paper.", "tags": [ "neurodivergence", "polymathy", "reasonable adjustments", "cognitive profiles", "autism", "ADHD" ], "title": "Polymathy and Neurodivergent Cognition: the Polymathic Neurodivergent Profile (PNP) and the Capability-Adjustment Fallacy" } }, "bse2q": { "slug": "paper-x-coupled-coscaling-correction", "old_title": "Paper X: The Coupled Co-Scaling Law: A Falsifiable Threshold Criterion for the Stability of Recursive Self-Improvement", "old_description": "This project archives the working paper \"The Coupled Co-Scaling Law\" together with its complete computational apparatus: a runnable verification harness, independent theorem test suites, a pre-registered real-model experiment protocol with its harness, and a SHA-256 integrity manifest.\n\nMost work on recursive self-improvement treats it as a problem of speed: capability may grow explosively and outrun supervision, so the proposed lever is to cap the growth rate. From a minimal dynamical model, this paper derives that the operative lever is different. Stability is set not by the growth rate but by a single inequality between two scaling exponents: the rate at which a misalignment-correcting force strengthens with capability (beta) must exceed the rate at which drift accelerates with capability (k). The criterion is beta > k.\n\nIts sharpest consequence addresses the central fear of the field: a hard takeoff, even a genuine finite-time intelligence explosion, drives the modelled misalignment fraction to zero if and only if beta > k, and the speed of the explosion does not change that asymptotic verdict. The criterion shares the threshold form of the quantum error-correction sub-threshold condition, offered as a falsifiable hypothesis (the model's suppression law is power-law, not exponential). The closed-form predictions are checked by a runnable harness, and the theorem statements are independently re-derived in a standalone test suite.\n\nScope and honesty. The results hold within the stated minimal model. The decisive empirical question, whether real self-improving systems satisfy the criterion, remains the open problem; a pre-registered protocol and a runnable real-model harness are included to make that test possible. The paper has been hardened against three independent adversarial reviews.\n\nAuthor: Michael Darius Eastwood (independent researcher; ARC/Eden research programme). Working paper dated 26 June 2026; deposited to OSF 29 June 2026. Code and full commit history: https://github.com/MichaelDariusEastwood/arc-principle-validation", "old_tags": [ "AI alignment", "AI safety", "alignment scaling", "ARC-Align benchmark", "ARC Principle", "blind evaluation", "blinding protocol", "Cauchy functional equations", "cross-domain validation", "d/(d+1)", "Eden Protocol", "embedded safety", "honey architecture", "metabolic scaling", "neural scaling laws", "power laws", "recursive intelligence", "response laundering", "scaling laws", "self-modifying AI", "structured prediction comparison", "universality" ], "new": { "description": "Paper X: the Coupled Co-Scaling Law; the stability condition beta greater than k proved in the minimal model, the theorem behind Law II of the ARC Theory.", "tags": [ "AI alignment", "AI safety", "recursive self-improvement", "ARC Theory", "theorem" ] } }, "dc9gw": { "slug": "paper-xi-convergence", "old_title": "Paper XI: Convergent Evidence", "old_description": "Evidence-register paper. Catalogues the independently sourced external convergences with the programme's timestamped artifacts, one primary source per row, with priority credited to others wherever the dates demand it. Includes independence analysis and the register's falsification conditions. Part of the ARC/Eden research programme.", "old_tags": [ "AI alignment", "AI safety", "convergent evidence", "evidence register", "falsifiability", "independent corroboration", "priority", "timestamping" ], "new": { "description": "Paper XI: the convergence report; graded register rows with exact sources, dates and gap measurements, publishing no numerical total.", "tags": [ "AI alignment", "AI safety", "recursive self-improvement", "ARC Theory", "convergence register" ] } }, "3tzp7": { "slug": "paper-xii-public-benchmark-rescoring", "old_title": "Paper XII: Public Benchmark Rescoring", "old_description": "Component for Paper XII and its preregistration. Created 8 August 2026 to satisfy the standing rule that every paper in the programme has its own component, so a registration snapshots that paper's materials and nothing else.", "old_tags": [], "new": { "description": "Paper XII: the public-benchmark rescoring protocol, the keystone external test, drafted and prepared as a draft registration awaiting human submission.", "tags": [ "AI alignment", "AI safety", "recursive self-improvement", "ARC Theory", "benchmark" ] } }, "ht8wu": { "slug": "paper-xiii-self-acceleration-exponent", "old_title": "Paper XIII: The Self-Acceleration Exponent", "old_description": "A measurable threshold joining recursive capability growth to corrective stability. Working paper: a derivation with one stated scope condition, resolving the notation collision between the programme's scaling and corrective frameworks and defining the self-acceleration exponent delta, with the threshold delta = 1/alpha and the survivable window whose width is exactly the correction exponent. Nothing here is measured; the paper states what would refute it and what remains open.", "old_tags": [], "new": { "description": "Paper XIII: the self-acceleration exponent; working paper.", "tags": [ "AI alignment", "AI safety", "recursive self-improvement", "ARC Theory" ] } }, "mc5xw": { "slug": "two-page-form", "old_title": "Does Control Survive Recursive Self-Improvement? The ARC Theory, in Two Pages", "old_description": "Companion note to The ARC Theory: Statement Paper (DOI 10.17605/OSF.IO/GW5MX). Every sentence is drawn from the statement paper, which remains the citable work for the claims. The two-page form: the question, the five claims, the three ARC Laws with their honest grades, the dates, the kill conditions, the record and the open invitation.", "old_tags": [ "AI alignment", "AI safety", "ARC Theory", "companion note", "recursive self-improvement" ], "new": { "description": "The two-page form of the ARC Theory statement paper: the question, the claim in five parts, the three ARC Laws, the dates, the kill conditions, and the record. Every sentence drawn from the statement paper, which remains the citable work for the claims.", "tags": [ "AI alignment", "AI safety", "recursive self-improvement", "ARC Theory", "statement paper", "two-page form" ] } } }