Evidence-Gated Recursion: An Experiment on a Self-Governing Research Organization of Frontier Language-Model Agents

Papers written by the autonomous-research system FARS averaged 3.17 of 10 in human review, and Into the Unknown found integrity failures in 114 of 135 agent-written research submissions. We report an experiment on a self-governing research organization of frontier language-model agents that worked only on problems with a named state of the art and an exact or official scorer, which no agent’s judgment can move. Evidence-gated recursion (EGR), the organization’s governance protocol, gives each type of candidate one gate, a predicate over one kind of evidence, and the kinds range from exact certificates and official scores to preregistrations, cross-family audits and number traces. A candidate is admitted only when its gate accepts it, and the organization applied the same rule to its own instruments at four levels. The experiment produced a new upper bound C₁ ≤ 1.50266484… for the first autocorrelation inequality, below the Einstein Arena leader’s 1.5027434867, and circle-packing records at 306 of Packomania’s sizes. Both stood on October 30, 2026. With Claude Opus 5.5 as the worker model, its take-home harness reached a median of 914 cycles against 948 for a CORAL-faithful team. We also found that weak-to-strong memory transfer is negative and that verdict-based supervision lowers a frontier model’s production correctness. Cross-family audits failed 11 of 60 of the organization’s own claims. No number in 693 traced claims is untraceable. The headroom relation, found in a post-hoc audit of the 195-comparison record, states that a component’s lift is a share of the model’s residual error. Fitted on weaker models, the relation forecasts Claude Opus 5.5 out of sample.