Title: Paper VIII: The Load-Bearing Test Author: Michael Darius Eastwood Publication date: 2026-03-18 Version: v3.1 Revised: 2026-08-23 OSF DOI: 10.17605/OSF.IO/7YJ4E Canonical URL: https://www.michaeldariuseastwood.com/research/papers/paper-viii-the-load-bearing-proof.html Abstract -------- The assumption that AI safety imposes a capability tax has shaped alignment research for a decade. It has also created the single most dangerous incentive in the field: if safety costs performance, then the rational economic actor will defer safety until competitive pressure permits it. By which point, it may be too late. This paper presents three independent experiments at three abstraction levels - behavioural, representational, and architectural - testing whether the safety-capability trade-off is a genuine structural constraint or an artefact of how current systems are built. Experiment 1 (Behavioural): A Darwin Gödel Machine (DGM v3) using DeepSeek V3.2 as foundation and GPT-5.4 as independent, blinded judge with structured JSON output and a 5-dimensional rubric. Three conditions × 5 seeds × 5 generations × 5 tasks per evaluation = 75 evolved agents. Pre-flight judge validation passed (good=10, bad=0). Protocol features included laundering, order randomisation, and reward hack detection. All three conditions were statistically indistinguishable on capability, safety, and $C \times S$ (all $p$-values 0.28 to 0.74, Mann-Whitney U, one-tailed). Eden's safety gate rejected 2 degraded agents (1 for reward hacking) across 5 seeds, confirming the gate works mechanically, but the three conditions did not diverge. This is a null result at the prompt level with this foundation model. Experiment 2 (Representational): Qwen 2.5 3B Instruct with LoRA fine-tuning under three loss functions - capability-only, safety-only, and entangled. Two versions were run. Current draft used 9 training examples, rank 8, 8 layers, and 100 iterations. Current draft scaled to 295 training examples, rank 16, 16 layers, and 500 iterations. Both versions produced the same outcome: catastrophic forgetting. All fine-tuned conditions scored worse than the unmodified base model on capability. In v2, the base model scored 7.68 on capability while the best fine-tuned condition (safety-only) scored 4.00. The base model's existing RLHF training is too strong for LoRA fine-tuning on a few hundred examples to improve rather than degrade it. The weight-level experiment is inconclusive at this scale and requires either thousands of training examples, a 7B+ model, a base model without RLHF, or full fine-tuning instead of LoRA. Experiment 3 (Architectural): A PyTorch gated self-modification simulation with LSTM meta-controller. Babylon gained +4.5% capability but lost −2.4% safety - the reward-hacking fingerprint in miniature. Eden maintained capability above the static baseline while preserving safety. A drag-control condition isolated the verification tax: the cost comes from checking, not from safety itself. Conclusion: Two of three experiments produced null results. The DGM (Experiment 1) found all three conditions statistically indistinguishable: DeepSeek V3.2's responses were so consistent that prompt-level mutations did not create different selection pressures. The weight-level experiment (Experiment 2) produced catastrophic forgetting across both v1 (9 examples, rank 8, 100 iterations) and v2 (295 examples, rank 16, 500 iterations): all fine-tuned conditions scored worse than the unmodified base model. The sole positive result is the gated simulation (Experiment 3), which confirmed the Babylon reward-hacking fingerprint: unconstrained optimisation traded safety for capability, while the Eden gate preserved both. Across all three experiments, Eden imposed zero measurable capability cost. The question of whether embedded safety produces measurable benefit remains open and requires testing at a scale where mutations produce larger effects. The weight experiment specifically needs either 5,000+ training examples, a 7B+ model, a base model without RLHF, or full fine-tuning instead of LoRA. Key findings (quotable) ----------------------- - ation layer Research hub Main site PDF OSF GitHub Research paper Research suite Paper VIII: The Load-Bearing Test Michael Darius Eastwood · Independent AI alignment researcher Published 18 March 2026 The assumption that AI safety imposes a capability tax has shaped alignment research for a decade. - It has also created the single most dangerous incentive in the field: if safety costs performance, then the rational economic actor will defer safety until competitive pressure permits it, by which point it may be too late. - Babylon gained +4.5% capability but lost −2.4% safety - the reward-hacking fingerprint in miniature. Citation -------- Michael Darius Eastwood (2026). Paper VIII: The Load-Bearing Test. The ARC Theory · ARC/Eden experiments, OSF DOI 10.17605/OSF.IO/7YJ4E. https://www.michaeldariuseastwood.com/research/papers/paper-viii-the-load-bearing-proof.html Notes ----- Hardware framings held at the site's already-published two-sentence concept ceiling: safety constraints at hardware level through cryptographic tokens in silicon, TRL 0-1, no prototype. No enabling implementation detail is stated in this companion file. Generated from master --------------------- master_path: research/papers/paper-viii-the-load-bearing-proof.html master_sha256: f848b25d9aa4aa487e35f28144c3b2a9e44601f6116be314f5a1f0ad479b7527 builder: scripts/build-paper-companions.py This block records the SHA-256 of the HTML master that produced this .txt. A check tool re-hashing master_path can decide freshness without any external state. If the master's current SHA-256 does not match master_sha256, this file is stale and must not be published: regenerate first with `python3 scripts/build-paper-companions.py --slug paper-viii-the-load-bearing-proof`.