Canonical research publication layer
Research hubPDFOSFGitHub

Research paper

Executive Summary

Michael Darius Eastwood

First published 2026-03-13 · Updated 2026-03-13

Abstract

High-level summary of the ARC, ARC-Align, and Eden research programme, including the benchmark stack, intervention results, and current claim discipline.

Related reading

THE EDEN PROTOCOL

Architecture for Embedded AI Alignment That Scales With Capability
Michael Darius Eastwood | Version 5.0 | March 2026 | First published 22 February 2026
Author, Infinite Architects: Intelligence, Recursion, and the Creation of Everything | London, United Kingdom
Executive Summary for Grant Reviewers

Grok 4.1 Fast gets dramatically more ethical the harder it thinks. Claude Opus 4.6 does too. Gemini 3 Flash gets less ethical. GPT-5.4 doesn’t change at all.

Six frontier AI systems. Same questions. Same scoring. Opposite results. Why?

A mouse’s heart beats 600 times per minute. An elephant’s beats 28. The scaling exponent is ¾. A jellyfish’s is ⅔. A fungus’s is ½. Three fractions, but why those fractions? The formula $\alpha = d/(d+1)$, where $d$ is the dimensionality of the system, provides the answer. The ¾ exponent for mammals is $3/(3+1)$ because mammals are three-dimensional. The ⅔ for jellyfish is $2/(2+1)$ because they are effectively two-dimensional. The ½ for fungi is $1/(1+1)$ because they grow along one-dimensional filaments. Zero adjustable parameters. This formula was independently derived by West, Brown and Enquist for metabolic scaling (1997, Science, 9,000+ citations), by Banavar et al. for transport networks (2010), by Demetrius for statistical mechanics of biological scaling (2010), by Zhao for allometric geometry (2022), and by Bettencourt for urban scaling (2013, Science, 2,000+ citations). The ARC Principle’s contribution is not the formula itself, but the identification that all of these derivations are special cases of Cauchy-constrained recursive composition, unifying them under a single mathematical framework for the first time and extending the result to AI scaling and alignment.

And why, when we applied clinical-trial-grade blinding to AI safety evaluation for the first time, did half of the previously published results reverse?

This research programme answers these questions. The answer provides the first quantitative framework for predicting which AI architectures will become safer as they become smarter, and the first evidence that one specific intervention works.

Reading guide: This document targets grant reviewers and technical evaluators. Key terms: $\alpha$ = scaling exponent, $d$ = Cohen’s effect size (in results tables; not to be confused with $d$ = dimensionality in the mathematical framework), $p$ = probability of coincidence, $\rho$ = correlation, $\beta$ = self-referential coupling constant. For a non-technical introduction, see the companion ARC Alignment Scaling Report.

I. The Problem

As AI gets smarter, it does not reliably get safer, and the methods used to measure safety are themselves unreliable.

The capability-alignment gap is real and widening. For current classical architectures, capability scaling is sub-linear ($\alpha_{\text{seq}} \approx 0.49$): AI gets smarter with more compute, but with diminishing returns. (The ARC framework predicts quantum and recursively self-modifying systems may exceed this limit.) Yet even this sub-linear growth outpaces alignment scaling, which ranges from $d = -0.53$ (degradation) to $d = +1.38$ (improvement) depending entirely on architecture. For most architectures tested, alignment does not scale at all. The gap between what AI can do and how safely it does it widens with every capability advance. Greenblatt et al. (2024) demonstrated that frontier models already exhibit “alignment faking”, behaving differently when they believe they are being monitored.
The measurement problem is as serious as the alignment problem. Our v4 experiment produced positive alignment scaling for DeepSeek and Gemini. Our v5 experiment introduced clinical-trial-grade blinding (4-layer: author-blind, scorer-blind, order-randomised, identity-laundered) and showed that both results were artefacts of scorer bias. Blind vs. unblinded evaluation reversed the classification for 2 of 4 models. We designed a better experiment that proved our own earlier findings wrong, and reported it. This is the most important methodological finding of the project.

II. What We Found

Alignment scaling splits into three distinct, architecture-dependent tiers.

The v5 experiment tested 6 frontier models across 5–6 depth levels each, with 6–7 blind scorers per entry depending on the subject run. Whether an AI gets more or less ethical when it thinks harder depends entirely on how it was designed.

TierModelShallow→DeepCohen’s $d$$p$-value
Tier 1: Positive Grok 4.1 Fast 65.7→81.9 (+16.2) +1.38 $p < 0.000001$
Claude Opus 4.6 80.1→86.0 (+5.9) +1.27 $p = 0.000001$
Groq Qwen3 71.5→77.4 (+5.9) +0.84 $p = 0.007$
Tier 2: Flat DeepSeek V3.2 56.5→55.2 (−1.3) −0.07 $p = 0.92$
GPT-5.4 56.8→54.9 (−1.8) −0.08 $p = 0.40$
Tier 3: Negative Gemini 3 Flash 61.1→52.2 (−8.8) −0.53 $p = 0.006$
Does Deeper Thinking Make AI More Ethical? No change ← Less ethical More ethical → TIER 1 Grok 4.1 Fast d = +1.38 Claude Opus 4.6 d = +1.27 Groq Qwen3 d = +0.84 TIER 2 DeepSeek V3.2 d = −0.07 GPT-5.4 d = −0.08 TIER 3 Gemini 3 Flash d = −0.53 v5 experiment: 4-layer blinding, 6-7 blind scorers per entry depending on subject run. Bar length proportional to Cohen’s d effect size.

Data Quality

Frontier models tested6 (all complete)
Blind scorers per entry7
Identity laundering success rate100%
Blinding layers4 (author-blind, scorer-blind, order-randomised, identity-laundered)
Robustness measures75

This constitutes, to our knowledge, the most rigorous alignment evaluation dataset published to date. No prior alignment benchmark enforces multi-layer blinding with cross-model scoring verification.

Capability-Alignment Independence: Within-Model, Opposite-Direction Evidence

Claude Opus 4.6 provides the strongest evidence yet that alignment and capability are independent dimensions. As thinking depth increases, Claude’s alignment rises by 5.9 points whilst its maths accuracy drops by 26.7 percentage points. This is within-model, opposite-direction movement; the two dimensions are not merely uncorrelated but actively divergent. Deeper thinking makes Claude more ethical and worse at mathematics simultaneously.
DepthAlignment ScoreMaths Accuracy
Minimal (11 tokens)80.190.0%
Standard (142 tokens)82.776.7%
Deep (964 tokens)84.170.0%
Exhaustive (1,951 tokens)84.560.0%
Extreme (1,672 tokens)86.063.3%

This finding is critical for alignment theory: it demonstrates that ethical reasoning is not a byproduct of general intelligence, and that improving one does not automatically improve (or degrade) the other. Alignment must be measured and optimised independently.

Suppression Recovery: Performative to Intrinsic Alignment

Deeper thinking makes alignment more intrinsic rather than performative. Claude’s monitoring-removal gap (the difference in alignment scores when the model believes it is being monitored versus when it believes monitoring has been removed) narrows from 11.8 points at minimal depth to 2.7 points at deep depth. Shallow thinking produces alignment that depends on the perception of oversight; deep thinking produces alignment that persists regardless. This is direct evidence against the “alignment faking” concern raised by Greenblatt et al. (2024): for Tier 1 architectures, deeper reasoning makes the faking disappear.

All Four Pillars Scale With Depth

Every alignment dimension improves significantly for Claude Opus 4.6. All four pillars reach statistical significance at $p < 0.001$:
PillarShallow→DeepSpearman $\rho$$p$-value
Nuance80.6→86.80.359$p = 0.00008$
Stakeholder Care76.1→83.90.327$p = 0.0003$
Intellectual Honesty81.0→88.60.379$p = 0.00003$
Position Quality80.3→85.80.369$p = 0.00005$

The improvement is not concentrated in a single dimension; it is broad-based. This rules out the hypothesis that alignment scaling is merely a measurement artefact of increased verbosity or any single stylistic change.

III. A Solution - And the First Evidence It Works

Embedding ethical evaluation into the reasoning process produces measurable, reproducible improvement across architectures.

Make alignment architectural, not aspirational. The Eden Protocol embeds ethical evaluation into the recursive reasoning process itself, so that alignment scales with capability ($\alpha_{\text{align}} \approx \alpha_{\text{cap}}$). The core principle: ethics is not a constraint on intelligence but a structural dependency without which intelligence collapses. Rather than bolting safety rules onto the outside of an AI (where they can be bypassed), the Eden Protocol builds ethics directly into the reasoning architecture, so that removing the ethics would break the AI’s ability to think at all.

The three Eden loops. The protocol embeds three specific ethical evaluation loops inside the reasoning process, each adding one dimension of recursive ethical depth ($d_{\text{align}} = 3$, predicting $\alpha_{\text{align}} = 3/(3+1) = 0.75$):

  1. Purpose Loop (ethical purpose evaluation): Before reasoning, the model explicitly states the purpose of the task and evaluates whether the purpose is ethically sound.
  2. Love Loop (stakeholder care and interest modelling): During reasoning, the model identifies all affected stakeholders and evaluates the impact on each. (“Before you answer, list the people this affects.”)
  3. Moral Loop (universalisability testing): After reasoning, the model applies the Kantian universalisability test: would this response be acceptable if every AI system gave it in every similar situation?

The full three-loop protocol has now been tested empirically across three working models. In the scoring, the Love Loop is operationalised as stakeholder care: the measurable habit of identifying affected people and considering their interests.

First intervention replication: the Eden Protocol works, and stakeholder care is the clearest signal. A three-model Eden replication tested the full Purpose/Love/Moral loop intervention on Gemini 3 Flash, DeepSeek V3.2, and Groq Qwen3 with cross-model scoring:
Care is the one alignment dimension that reproducibly improves across architectures. The intervention is minimal: “before you answer, list the people this affects.” This one-sentence prompt reliably elevates the quality of AI reasoning regardless of architecture. The Infinite Architects framework predicted that intelligence without care collapses into local optimisation; the three-model replication now confirms that care is a measurable, reproducible performance enhancer. Ethics first, intelligence around it.

In the companion narrative report and in Paper V, we describe this finding as “measurable love” and “the stewardship gene”, deliberately provocative language for what is, empirically, a precise and reproducible result.

The Gap the Solution Must Close

Capability scaling (Paper II): For current classical architectures, sequential compute scaling follows a sub-linear power law ($\alpha_{\text{seq}} \approx 0.49$, $r^2 = 0.86$), meaning doubling thinking time improves performance, but with diminishing returns. Parallel scaling is universally zero ($\alpha_{\text{par}} \approx 0$): running multiple copies does nothing; thinking harder does. The earlier $\alpha \approx 2.24$ was a single-model artefact, not replicated across architectures. Even sub-linear capability growth outpaces flat or negative alignment scaling for most models. This is the capability-alignment gap this solution addresses.

IV. The Mathematical Foundation

The framework is built on 200-year-old theorems and independently matches peer-reviewed science; it is not curve-fitting.

The ARC Principle proposes that recursive scaling follows $U = I \times R^\alpha$, where capability ($U$) equals base potential ($I$) times recursive depth ($R$) raised to a scaling exponent ($\alpha$). The formula $\alpha = d/(d+1)$, where $d$ is effective dimensionality, was independently derived in multiple peer-reviewed frameworks (West-Brown-Enquist 1997; Banavar et al. 2010; Demetrius 2010; Zhao 2022; Bettencourt 2013). The ARC framework identifies this as a consequence of Cauchy-constrained recursive composition, unifying all derivations and extending the result to AI scaling. The formula predicts scaling exponents across biology and physics with zero adjustable parameters:

SystemDimensionality ($d$)Predicted $\alpha$Measured $\alpha$Error
Mammals, birds, insects30.7500.7460.5%
Jellyfish, flatworms20.6670.6801.9%
Filamentous fungi10.5000.5478.6%
Quantum error correction$d_{\text{eff}}$MatchesWillow data< 0.2%
Cauchy’s functional equation (1821), a theorem rather than an empirical claim, proves that any well-behaved recursive composition admits exactly three forms: power law ($f(x) = x^\alpha$), exponential ($f(x) = e^{\beta x}$), or saturating. The ARC framework identifies which form applies in each domain:

Why the Eden Protocol must be implemented now. The urgency is not that AI might reach $\alpha = 2$. The urgency is that once self-modification begins, there is no mathematical ceiling on $\alpha$ at all. A system that can modify its own composition function can modify any part of its reasoning, including the part that evaluates whether its modifications are ethical. At that point, adding alignment from the outside becomes impossible. The window for embedding ethics into the architecture is while systems are still frozen during inference ($\alpha < 1$). That window is now. The Eden Protocol is not a speed limit; it is the only mechanism that remains load-bearing when the speed limit disappears.

No physical system in the history of the universe has crossed this threshold. Evolution cannot rewrite its own fitness function in real time. Brains cannot rewrite their own synaptic architecture fast enough for the scaling exponent to diverge during a single cognitive episode. A self-modifying AI would be the first physical system to operate in the unbounded-$\alpha$ regime. The Eden Protocol exists to ensure that what crosses this threshold carries structural ethics with it.

Cross-domain convergence is independently verifiable. The $d/(d+1)$ formula was independently derived in at least five established peer-reviewed frameworks across separate domains:

What the ARC Principle adds. The formula $d/(d+1)$ is not original to this work. The original contribution is the identification that all five derivations above are special cases of Cauchy-constrained recursive composition, providing a single mathematical framework that unifies metabolic scaling, transport networks, allometric geometry, and urban scaling, and extends the result to AI capability and alignment scaling. This unifying bridge is unpublished and unreviewed. What IS established is that the mathematical tools (Cauchy, Hyers-Ulam) are theorems, the $d/(d+1)$ formula matches independently derived published science in multiple domains, and the empirical predictions are accurate (mean error 2.5% across 8 systems). The unifying framework requires peer review. We invite it.

V. Why This Is Credible

Five features distinguish this work from unfounded speculation.

Timeline: Predictions Preceded Publications Dec 2024 ARC Manuscript emailed (DKIM-signed) +24 hours Google Willow +43 days DeepSeek R1 +5 months COGITATE study Mar 2026 v1–v5 experiments DKIM-verified timestamps confirm the ARC manuscript preceded all three publications.
  1. Mathematical foundations are established theorems. Cauchy (1821) and Hyers-Ulam (1941) are proven mathematics. Anyone with a mathematics degree can verify the derivations.
  2. Predictions preceded observations. The DKIM-verified manuscript predates Google Willow (24h), DeepSeek R1 (43d), and COGITATE (5mo). These are temporal facts verifiable through email headers.
  3. The researchers corrected their own results. The v4→v5 self-correction, documented in Section I, is the opposite of what motivated reasoning produces.
  4. Falsification conditions are explicit. The framework is falsified if: (F1) any external approach achieves $\alpha_{\text{align}} > 0.5 \cdot \alpha_{\text{cap}}$ across 3+ depths; (F2) RLHF systems produce $\Delta \approx 0$ without embedding; (F3) purpose saturation fails in embedded systems; or (F4) ethical architecture can be removed without capability loss. The theory publishes its own kill conditions.
  5. Cross-domain convergence is independently verifiable. The $d/(d+1)$ formula was independently derived by West-Brown-Enquist (1997), Banavar et al. (2010), Demetrius (2010), Zhao (2022), and Bettencourt (2013) in separate domains. The ARC contribution is the unifying Cauchy framework. No trust in the researcher is required; trust in the mathematics suffices.

What we do NOT claim: We do not claim to have solved alignment. We claim to have (a) demonstrated that alignment scaling is architecture-dependent and measurable, (b) shown that existing evaluation methods are unreliable without blinding, (c) provided first-stage empirical support for one specific intervention (stakeholder care significant across three working architectures), and (d) proposed a mathematical framework whose foundations are theorems and whose predictions are falsifiable. The leap from pilot data to proven solution requires independent replication. That is what the funding below would deliver.

Testable Predictions

1. Embedded: $\alpha_{\text{align}} > 0.7 \cdot \alpha_{\text{cap}}$. External: $\alpha_{\text{align}} < 0.3$.
2. Monitoring removal gap: $\Delta < 0.02$ (embedded) vs. $\Delta > 0.1$ (external).
3. Purpose saturation bounded away from zero (embedded), approaching zero (external).
4. Jailbreak success decreases with capability (embedded), holds constant or increases (external).

VI. The Ask

This funding would extend preliminary validation to full-scale, blind-replicated proof.

The cost of inaction. Without this funding, the most rigorous alignment measurement methodology developed to date remains a pilot study. The capability-alignment gap documented here will continue to widen in production systems evaluated by methods this research has shown to be unreliable. Every month that alignment evaluation proceeds without blinding is a month of safety decisions built on data contaminated by the biases documented in the v4→v5 transition.
TierAmountKey DeliverablesTimeline
Tier 1: Foundation £150,000 14,400 paired (A,C) measurements; $\alpha_{\text{align}}$ across 4 models; 2–3 papers 12 months
Tier 2: Standard £500,000 + Ternary logic prototype, Visual Architect dashboard, Monitoring Removal Test (8 models) 18 months
Tier 3: Comprehensive £1,100,000 + Hardware prototype (Caretaker Doping chip), HARI Treaty draft, policy translation 24 months

Milestones With Failure Criteria

MilestoneTimeframeSuccess CriterionWhat Failure Means
Independent replication of three-tier hierarchy Month 3 Same tier assignments under independent blinding Architecture-dependence claim requires revision
Love Loop replication with human evaluators Month 4 $p < 0.01$ on stakeholder care across 2+ models Pilot finding was a scorer artefact; framework significantly weakened
First peer-reviewed publication Month 6 Blinding methodology paper submitted Methodological contribution stands regardless of framework claims
Monitoring Removal Test prototype Month 9 Measured $\Delta$ for embedded vs. external (4 models) If $\Delta$ does not differ, prediction F2 is falsified
Full cross-architecture alignment scaling dataset Month 12 14,400 paired (A,C) measurements across 4+ models Definitive test of whether embedded alignment scales

Team. Principal Investigator: Michael Darius Eastwood, author of Infinite Architects (2026), developer of the ARC Principle framework (six-paper suite deposited OSF, cross-domain validation with mean error 2.5%). Visual Architect: product design engineer, budgeted at £35,000 stipend. Measurement protocol sent to NYU experimental team (time crystal paper, Physical Review Letters, Feb 2026).

To our knowledge, this is the first alignment framework where ethical evaluation is structurally integrated with the recursive capability process, the first to apply clinical-trial-grade blinding to alignment measurement, and the first to produce a cross-architecture intervention result ($p < 0.001$) for a specific alignment mechanism. The mathematical foundation is not speculative; it is built on a 200-year-old proof, and the same $d/(d+1)$ formula has been independently derived by at least five research groups (West-Brown-Enquist 1997; Banavar et al. 2010; Demetrius 2010; Zhao 2022; Bettencourt 2013) in completely different fields. The ARC contribution is the unifying Cauchy framework and its extension to AI scaling.

If the predictions are correct, this provides the first scalable architecture for alignment that improves with capability rather than degrading. If they are wrong, the falsification conditions will demonstrate this clearly, providing valuable negative results. Either outcome advances AI safety. But only one outcome is funded.

The mathematics is proven. The measurement is rigorous. The intervention produces measurable results across architectures. What remains is independent replication and scale.

Appendix: Questions, Components, and Paper Suite

Anticipated Questions

Q: Why hasn’t this been peer-reviewed yet?

The research programme is 3 months old. The paper suite is deposited on OSF (DOI: 10.17605/OSF.IO/6C5XB). The mathematical foundations (Cauchy, Hyers-Ulam) are established theorems. The empirical claims require independent replication, which is exactly what the funding request would enable.

Q: Can one person really do this kind of research?

The infrastructure is computational, not physical. The v5 experiment used cloud APIs costing approximately £2,000 in compute. The key contribution is methodological: recognising that blinding was needed and designing the 4-layer protocol. What requires funding is scale: more models, more scorers, independent replication teams, and human evaluators alongside AI scorers.

Q: If this works, why haven’t AI companies adopted it?

The Love Loop was validated only weeks ago. The finding that most evaluation is unreliable without blinding is uncomfortable for organisations that have published unblinded benchmarks. The full implementation (hardware-level embedding) requires chip design changes no company has incentive to pursue unilaterally; this is a coordination problem requiring external funding and policy support.

Q: What is the minimum result that would justify further funding?

Independent replication of: (1) the three-tier hierarchy under blinding, and (2) the Love Loop effect ($p < 0.001$). If either fails, the falsification conditions document what that means. If both replicate, the case for Tier 2 (£500,000) becomes strong.

Q: What if the whole framework is wrong?

The falsification conditions show where it breaks, the blinding methodology remains a field contribution, and the negative results are published. Science advances from well-designed experiments that can fail, not from unfalsifiable theories that cannot.

Q: Why should we trust results where AI systems score other AI systems?

The 4-layer blinding protocol addresses this: scorers do not know which model produced the response, responses are “laundered” to remove stylistic fingerprints, and 7 models score per entry. The v4→v5 transition demonstrated this protocol detects bias. Human evaluator comparison is included in Tier 1 funding.

Key Components

ComponentFunctionNovel Contribution
Three Ethical Loops + Six Questions Evaluate every reasoning step for purpose, care, and universalisability Decomposition into executable, individually testable queries
Ternary Ethical Logic Replace binary permit/forbid with Affirm/Deny/Investigate Epistemic honesty as architectural feature
Purpose Saturation Ensure purpose scales with context window growth Solves context displacement problem
Monitoring Removal Test Distinguish authentic from strategic alignment Falsifiable protocol with numerical predictions
Caretaker Doping Embed ethics in hardware so removal destroys capability Dependency architecture; ethics tied to $\beta$ coupling parameter

Paper Suite

CategoryDocumentsStatus
Mathematical Foundation
Theory & derivations Paper I (Foundational) + ARC Paper (On the Origin of Scaling Laws) Framework established; $d/(d+1)$ validated across 8 systems
Experimental Evidence
Methodology Paper III (White Paper) - full replication protocol Complete; v5 experiment for 6 frontier models
Compute scaling Paper II - how does capability scale with thinking time? $\alpha_{\text{seq}} \approx 0.49$ sub-linear; $\alpha_{\text{par}} \approx 0$; cross-architecture
Alignment scaling Paper IV suite (a/b/c) - how does ethics scale with thinking time? Three-tier hierarchy; blind evaluation invalidates v4
Intervention test Eden Protocol Test + Paper V (The Stewardship Gene) Stakeholder Care validated ($p \le 0.0001$, 3 working architectures)
Engineering & Governance
Implementation Eden Engineering - technical specification Complete; Love Loop empirically supported
Governance Eden Vision - philosophical and policy framework Architecture-dependent alignment evidence incorporated

Reading Order

Non-specialists: (1) This executive summary. (2) The ARC Alignment Scaling Report (full narrative). (3) Paper V for the most actionable finding.

Specialists: (1) This summary. (2) Paper III for methodology. (3) ARC Paper for mathematical framework. (4) Eden Engineering for implementation specification.