Why AI Self-Improvement Claims Need Falsifiable Gates, Not Self-Reported Wins

A developer built a meta-science framework for the All Things Agentic Hackathon to rigorously test whether AI self-improvement claims are genuinely verifiable. The core principle holds that no agent should be allowed to judge its own outputs, drawing on Popper's falsifiability standard and Pearl's distinction between correlation and causal intervention. Proposals from the AI model are placed in a non-authoritative tier and can only be promoted to canon by outperforming a baseline on unseen test worlds by a defined margin, preventing noise from being mistaken for progress. In three live runs, the system promoted only one of Gemini's three proposals, with most rejections involving real but insufficient gains. The project also turned its scrutiny inward, uncovering a flaw in its own benchmark methodology that was only caught through rigorous self-auditing.