Each model receives the others’ answers in its system prompt and is required to fact-check them. Model N reads and criticises models 1..N-1.
Nousaxis is a multi-agent workspace that runs four frontier AI models against the same question together. The models see and criticise each other’s answers, a judge evaluates all of them at once, and a verifier checks claims against live web evidence. The goal is not to make one model smarter — it is to make fabrication rarer.
Each model receives the others’ answers in its system prompt and is required to fact-check them. Model N reads and criticises models 1..N-1.
A single judge evaluates every output together, produces a weighted consensus, and flags disagreement when the models contradict each other.
A verifier layer extracts atomic claims and scores hallucination risk by checking them against live web evidence.
We measured whether cross-examination actually helps, on the live production pipeline, with a blind grader. The results — including the limitations and confidence intervals — are published in full.
Read the measurement report →Nousaxis is developed by Talonaris Technology, an applied-AI company. The company story, R&D line, vision and team live on the corporate site.
talonaris.com →