This preprint reports results from 4,680 controlled evaluations across six frontier AI models, four experimental scenarios, six mathematical reasoning domains, and 90 contamination-minimised, formally verifiable problems. The central finding is an architectural phase transition in AI reliability: above approximately 95% accuracy, error structure undergoes a qualitative shift from stochastic to systematic, at which point compute scaling becomes fundamentally ineffective. The Generator–Auditor–Adversary–Synthesizer (GAAS) architecture breaks the single-agent ceiling: single-agent inference plateaus at 93.0%; self-consistency scaling yields only +1.5 pp (p=0.317, NS); role-separated GAAS achieves 98.7% (p<0.001); role-specialised model diversity achieves 100% on this evaluation set (Wilson CI: 95.9–100%). This is Empirical Study 1 of a planned three-study programme. Study 1 covers determinate domains (mathematical and logical reasoning). Studies 2 and 3 will extend to semi-determinate and indeterminate domains. Files included:- Main manuscript (PDF)- Online Appendix 1: All 90 evaluation questions, verified ground-truth answers, complete raw model responses for S1–S4, error taxonomy classifications, and experimental protocols (DOCX)
Pandit et al. (Sun,) studied this question.