We present a controlled multi-model experiment testing whether structurally distinct AI systems independently converge on identical answers to foundational questions about persistent systems — without any exposure to the theoretical framework under investigation. Seven models spanning four major architectures (Google Gemma, Meta Llama, Microsoft Phi, OpenAI GPT-OSS) were run in cold-start conditions across three question chains: a neutral structural probe (Group A), a domain-disguised equivalent (Group B), and an adversarial counter-induction chain (Group C). The primary finding is that 6 of 7 models independently derived a triadic minimum structure for persistent systems in Group A, while Group B produced a dispersed distribution (2/3/3/3/3/3/4), ruling out question-framing as a driver of the A-group convergence. Group C's adversarial test yielded 7/7 models defending structural convergence even under direct counter-inductive pressure. A genuine 4:3 split emerged on one open question (B27), confirming that the experimental design does not suppress divergence when divergence is structurally warranted. Four models independently constructed fully self-consistent alternative theoretical frameworks. We interpret these results as evidence that the triadic minimum structure reflects a property of the problem space itself, not of any particular training corpus or architectural bias.
Chen et al. (Fri,) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: