Rewriting the Structure of Its Own Improvement Mechanism: A Preregistered Causal Test of Bounded Recursive Self-Improvement

Abstract

A system that rewrites its own improvement mechanism meets common definitions of recursive self-improvement whether or not the rewriting helps. In five earlier preregistered tests of a Level-4 bounded self-improving program-repair system, the rewritten mechanism was never shown to beat a frozen one. Development measurements located why: the rewritable part had only about 1.7 percentage points (pp) to give, while the headroom lay in the form of the policy the mechanism learns, which it could not change. We made the mechanism's structure (how evidence becomes a policy, including the policy's form) a gene its own editing procedure can rewrite, offered among options known not to help, and preregistered a causal test. One of two campaigns found and adopted, by its own two-stage validation, both parts of the best development setting (a tree-ensemble policy form and a different budget split) and rejected a decoy three times; the other never sampled the useful option. On 24 fresh confirmatory seeds the evolved mechanisms beat the default by +2.57 pp (95% CI 0.91 to 4.17; 18 of 24 seeds; exact one-sided Wilcoxon p = 0.0028) and an evidence-free replay of the same proposal process by +3.52 pp (p = 0.0011), meeting every preregistered criterion. The gain came entirely from the campaign that found the structural rewrite (+5.70 pp, 12 of 12 seeds, against -0.57 pp) and was confined to the training families: +0.26 pp on four held-out families and +0.39 bugs (p = 0.52) on 164 real bugs. A post-hoc leave-families-out analysis shows that no policy form transfers more than about 5 pp to unseen families (ceiling 54%): the limit is what the system observes. Rewriting the improvement mechanism paid off causally for the first time in this programme, through a selected (not invented) structural change, in-distribution only.