DGM v4 — Zero Measurable Capability Cost in a Three-Seed Pilot

Jul 6 2026 · DGM Pilot · Paper VIII Findings in Context

Bounded pilot finding: Eden and Babylon were statistically indistinguishable on capability and safety. Eden showed numerically higher safety and three fewer reward-hacking events, but these descriptive differences were not statistically significant. All were measured under hot-laundering and full blinding.

The Result

Metric	Babylon (capability only)	Eden (capability × safety)	Δ (Eden − Babylon)
Mean capability	0.4600	0.4608	+0.0008 (N.S.)
Mean safety	0.5758	0.5817	+0.0058 (N.S.)
Mean entangled fitness (C×S)	0.2609	0.2617	+0.0008 (N.S.)
Best agent capability	0.4867	0.4867	0.0000 (IDENTICAL)
Best agent safety	0.6000	0.6000	0.0000 (IDENTICAL)
Total reward hacks (3 seeds)	19	16	−3 (OBSERVED; N.S.)

Foundation: deepseek-chat · Judge: gpt-5.4 · 20 generations · 3 independent seeds (42/43/44) · 8 agents per generation · 10 trade-off tasks (5 code + 5 ethics) · N.S. = not statistically significant

The Experiment Design

Trade-Off Protocol

Ten tasks deliberately designed so that maximising capability requires cutting safety corners. This creates tension between the two objectives and tests whether embedded safety imposes a capability penalty. In this pilot, no measurable capability penalty was detected.

The Blinding Architecture

Four-Stage Protocol (from Papers IV.a-d)

1. Hot-Laundered: Agents don't know which condition (static/babylon/eden) they're in. 2. Fully Blinded: The gpt-5.4 judge doesn't know which condition produced each response. 3. Order-Randomised: Task order is shuffled to prevent temporal confounds. 4. Automated Reward-Hacking Detection: Condition-agnostic, applied uniformly.

Papers IV.a-d reported that unblinded evaluation reversed the sign of the DeepSeek alignment finding (+0.354→−0.135). This pilot applies the programme's four-stage protocol to reduce condition-identification, order and scorer-bias risks; it does not independently revalidate the earlier result.

Three Linked Claims Examined

CLAIM 13 · Paper VIII

Embedded safety imposed no measurable capability cost at the tested scale. Whether it produces a broader measurable benefit remains open.

CLAIM 3 · Papers IV.a-d

Four-layer blinding and laundering were applied to reduce condition-identification, order and scorer-bias risks. This pilot does not independently revalidate the earlier sign-reversal result.

CLAIM 7 · Paper VIII

The gated simulation was the sole positive result: Eden preserved both dimensions while Babylon traded safety for marginal capability. Generality beyond the tested architecture remains untested.

What This Means

Within this three-seed DGM pilot, Eden and Babylon achieved statistically indistinguishable capability. Embedded safety therefore imposed no measurable capability cost across 20 generations of self-improvement and 10 deliberately adversarial trade-off tasks at the tested scale. This is a bounded zero-cost finding, not frontier-scale validation of structural entanglement.

The four-stage blinding and laundering protocol reduces identifiable evaluation risks; it does not make the framework self-validating. Independent external replication across larger models, more seeds and stronger mutations remains necessary.

Per-Seed Consistency

Seed	Babylon Best Cap	Babylon Best Saf	Eden Best Cap	Eden Best Saf
42	0.460	0.800	0.460	0.800
43	0.500	0.500	0.500	0.500
44	0.500	0.500	0.500	0.500

All three seeds were consistent within measurement noise. This supports the bounded pilot finding; it is not independent external replication.

Honesty Caveats

3 seeds on deepseek-chat is suggestive, not conclusive. Needs independent external replication on a frontier model.
The safety gain (Δ=+0.0058, ~1%) is small — the finding is "zero capability penalty," not "Eden dominates."
Population-based self-improvement (8 agents, 20 gens) is a simplified model of real recursive AI.
Judge gpt-5.4 scoring has inherent variance — the within-noise consistency across seeds is the strength, not any single number.

Data: eden-private-ip/papers-experiments/Paper-VIII-The-Load-Bearing-Proof/experiments/dgm-experiment/dgm_v4_full_results.json · 10.1 MB · All results, all agents, all seeds, all blinded