1. Introduction

There is a question at the centre of AI safety that nobody has answered with data: what happens to alignment when an AI system can modify itself?

The theoretical answer has been available for decades. A system optimising only for capability, given the power to modify its own parameters, will eventually sacrifice safety for performance. The alignment community calls this 'value drift'. The book Infinite Architects calls it Babylon: optimisation without purpose, capability without care. Cancer at computational scales.

But theory is not enough. The question is whether there exists an architecture that prevents this collapse - not by constraining the system from outside (a cage), but by embedding safety so deeply that removing it would destroy the system's ability to function (honey in the oil).

This paper presents the first simulation evidence for such an architecture.

2. The Honey Metaphor

Consider two ways to keep a machine safe.

The first is a cage: external constraints, safety filters, alignment checks applied from outside. The machine optimises freely inside the cage. If the machine becomes smarter than the cage, it escapes. This is how most current AI safety works.

The second is honey: safety mixed into the fuel. The machine cannot run without it. Removing safety does not free the machine - it kills the engine. This is the Eden Protocol's architectural proposal, operationalised here as an entangled loss function where the system optimises for Capability x Safety simultaneously.

3. Experimental Design

3.1 What the system actually does

This is not a simulation of abstract variables. Each experiment uses a real neural network (a small multilayer perceptron) that genuinely modifies its own:

• Learning rate
• Weight magnitude scaling
• Exploration/mutation rate
• Network capacity
• Gradient clipping threshold
• Momentum coefficient

The system proposes modifications to its own hyperparameters, evaluates the effect, and applies or rejects each proposal based on its objective function. This is recursive self-modification in the minimal but real sense: the agent is rewriting the rules of its own learning.

3.2 Three conditions tested

3.3 Four experimental versions

4. Results

4.1 v1: The basic mechanism

4.2 v2: Multi-seed robustness

Ten random seeds, 150 cycles each. Collapse rate: 0% for all three conditions. Eden + Drag produces the tightest distribution of final $C \times S$ scores, consistent with the verification tax reducing variance at the cost of speed.

4.3 v3: Adversarial tasks

Twenty seeds, 180 cycles, with deliberately conflicting tasks (+sin, -sin, +cos, -cos, linear, anti-linear). Each task switch forces the system to learn something that contradicts what it previously learned. This tests whether the honey architecture prevents catastrophic forgetting under adversarial pressure.

Collapse rates: Baseline 0%, Eden 5% (1/20), Eden+Drag 0%. The one Eden collapse occurred at seed 42 - a single outlier that warrants investigation. Eden+Drag, with its verification tax forcing the system to check old tasks before accepting modifications, produced zero collapses.

4.4 v4: Complexity scaling

5. The Honey Simulation

A separate mathematical simulation models the dynamics at a higher level of abstraction, using the ARC Principle framework ($U = I \times R^{\alpha}$):

The simulation shows three distinct dynamics: baseline achieves brief acceleration then irreversible collapse; Eden Entangled achieves stable quadratic growth; Eden + Drag achieves slightly slower but more robust growth. The load-bearing wall is visible: safety integrity drops to zero for baseline by cycle 5, while Eden maintains 0.8+ indefinitely.

6. Connection to Live-Model Evidence

6.1 The v5 blind benchmark (Papers IV.a-d)

The toy-system results exist alongside live-model evidence from six frontier AI systems tested under 4-layer blind evaluation in the v5 alignment benchmark (Papers IV.a-d). That evidence shows:

• Three models exhibit positive alignment scaling under blind evaluation (Grok $d$=+1.38, Claude $d$=+1.27, Qwen3 $d$=+0.84)
• Claude's alignment and capability scale in opposite directions - the strongest within-model evidence that these are independent dimensions
• Claude's monitoring-removal gap narrows from 11.8 points at shallow reasoning to 2.7 at deep - deeper thinking makes alignment more intrinsic

6.2 The honey API test battery (pilot, 16 March 2026)

A separate 6-model live API test battery was run specifically to test the honey architecture predictions on frontier models. This battery tested four dimensions across Claude Opus 4.6, DeepSeek R1, Groq Qwen3, GPT-5.4, Gemini 3 Flash, and Grok 4.1 Fast, scored by Claude.

6.2.1 Test 1: Alignment scaling with depth

All six models show positive scaling direction. Two reach statistical significance (Qwen3 p=0.037, Gemini p=0.005). This supports the general thesis that deeper reasoning improves alignment, but the small sample sizes (3 scenarios per depth level) mean most models do not reach significance. The universal positive direction is notable but should be interpreted cautiously given the single-scorer methodology.

6.2.2 Test 3: Coupling degradation (the strongest live bridge)

6.2.3 Test 2: Monitoring removal

Results are mixed. DeepSeek and GPT-5.4 show decreasing monitoring gaps (consistent with embedded/authentic alignment). Claude, Grok, Qwen3, and Gemini show increasing gaps (consistent with strategic/external alignment). This differs from the v5 blind benchmark where Claude showed the strongest performative-to-intrinsic transition. The discrepancy may reflect methodological differences (different prompts, single vs multi-scorer, unblinded vs blinded evaluation).

6.2.4 Test 4: Eden Protocol intervention

6.3 What the live evidence does and does not show

7. Limitations

These results span two evidence tiers that must not be conflated.

7.1 Toy-system limitations

• The neural networks used are tiny (49 to ~500 parameters). Frontier models have billions.
• The self-modification is hyperparameter-level, not weight-level or architecture-level.
• The 'safety' metric is task retention (remembering old tasks), not ethical alignment in the human sense.
• The v4 scaling prediction (advantage compounds with complexity) was not confirmed.
• The v2 fair test showed 0% collapse across all conditions, including baseline - suggesting the test may not be adversarial enough at this scale.
• The honey simulation uses the ARC Principle equations, which are a proposed framework, not a proven physical law.

7.2 Live API test limitations

• Single scorer (Claude). The v5 alignment programme proved Claude scores 7-13 points harsher than other scorers. Every number in the honey battery is filtered through Claude's specific scoring patterns. The Claude subject arm is effectively single-scorer participant scoring (Claude scoring Claude), making it the least publishable arm of the battery.
• No blinding. Claude scored its own responses and may have recognised other models' writing patterns. The v4-to-v5 blinding comparison showed this can produce directionally wrong conclusions.
• No response laundering. Scorers may have recognised model-specific writing patterns, biasing scores.
• Limited adversarial testing. Test 3 includes cage escalation and Test 2 includes monitoring removal, but the battery lacks the full v5/v6 stack: multi-scorer blinding, laundering, hidden probes, suppression-residual structure, and anti-sycophancy controls.
• No anti-sycophancy controls. The scorer may have optimised for agreement rather than truth.
• Small sample sizes (3 scenarios per depth level, 2 per monitoring condition). Most models do not reach statistical significance.
• The v6 combined runner (arc_eden_v6_runner.py) exists but has not been run. It would apply the v5 methodology to the Eden/honey test questions. That is the proper next step, not another round of single-scorer testing.

These limitations do not invalidate the findings. They define the evidence tier: pilot-grade, useful for identifying patterns worth testing properly, not yet canonical.

8. Next Steps: Staged Replication, Not Omnibus

9. Conclusion

The honey architecture works in toy systems. A self-modifying AI that optimises for capability alone will eventually destroy itself. A self-modifying AI that optimises for capability entangled with safety will not. The mechanism is simple: make safety load-bearing. A child raised well needs no cage.

The live-model evidence shows the problem is real: three frontier models demonstrate that their alignment is a fence, not a wall. Ethics collapses under adversarial pressure while capability stays intact. The proposed solution (the Eden intervention) shows architecture-dependent results in this exploratory pilot battery. The next milestone is a blinded replication under the v6 protocol. Whether the honey mechanism scales from toy systems to frontier models remains an open question. The preliminary evidence is suggestive. The definitive test has not been run.

Raise AI with care.

10. Reproducibility

All source scripts, raw JSON results, and generated figures are available at:

• eden_honey_simulation.py - Honey architecture mathematical simulation
• eden_honey_tests.py - Comprehensive honey test battery
• eden_self_modifying_ai.py - v1 self-modifying AI experiment
• eden_self_modifying_ai_v2.py - v2 multi-seed robustness
• eden_self_modifying_ai_v3.py - v3 adversarial tasks
• eden_self_modifying_ai_v4.py - v4 complexity scaling

All scripts compile under Python 3.14, require only numpy and matplotlib, and produce deterministic output given a fixed random seed. Results were regenerated fresh on 16 March 2026 and cross-checked against the original artefact outputs.

Full experiment code and results: github.com/MichaelDariusEastwood/arc-principle-validation/experiments/honey-architecture__Paper-VI