Pre-registered SLOPWARE E4: neutral-mutation accessibility predicts evolution at r=0.18
bushibuilds · x · 2026-10-14
SLOPWARE E4 is a pre-registered, run-once digital evolution experiment testing whether E3's accessibility findings predict evolutionary outcomes.
Key results
- Across 88 matched founder pairs (each program vs. a neutral descendant 60 bytes away with identical measured signatures), the rank correlation between predicted accessibility differences and observed mean final log₂ best fitness was 0.18 (95% CI −0.03 to +0.39) — null by the pre-registered rule, though the interval also covers the 0.3 deemed a substantial effect.
- Founder identity itself matters: behaviourally identical matched pairs reached different mean final fitness (S=0.47, permutation p=1/10001).
- Recomputed per generation, the correlation climbed to 0.43 around generation 31, significantly positive from generations 2–63, then fell back.
Setup: 16 populations of 64 per pair, 256 generations, one-byte mutations, selection purely on longest finite runs; generation-1's 180,224 proposals were checked against E3's map, generation 2+ unmapped. Author labels the outcome "Outcome D."
More from Research
- Correction Training frees diffusion models from CFG as looped architectures gain traction — CSProfKGD · 2026-10-14
- FramePack-U: Training-free constant-memory frame packing boosts long-video VLM benchmarks by up to 12.9% — UNC-ChapelHill · 2026-10-14
- CineRecon Benchmark: 12 Frontier Models Struggle to Reconstruct Editable 3D Scenes from Video — hkuhk · 2026-10-14
- Meituan's ClawLoop: Self-Improving Agents via Teacher-Free Harness-Native Training Loops — meituan-longcat · 2026-10-14
- SALD: Self-Referenced Advantage Learning Lets Diffusion Models Train Without a Teacher — IITK · 2026-10-14
- Scoring Higher, Answering Worse: ByteDance's ProRubric Fixes Reward Hacking in Rubric-RL — ByteDance · 2026-10-14