Structured Multi-Model Deliberation as a Method for Present-Day Frontier AI Capability Assessment

We report on a structured three-round deliberative assembly in which eight publicly accessible frontier language model deployments independently analyzed a shared technical motion concerning AI-assisted detection of human cognitive agency degradation. Rather than treating the deliberation as a mechanism for reaching consensus on the motion itself, we analyze it as an instrument for eliciting and comparing observable model capabilities: reasoning architecture, tool integration depth, epistemic honesty under uncertainty, falsifiability commitment, and novel signal generation. Across three rounds and one cross-pollination phase, we identify systematic capability differentials not captured by existing benchmark-based evaluations. We propose **structured deliberation** — in which models respond independently before convergence pressure is applied — as a complementary methodology to benchmark suites for present-day capability mapping. Our principal finding is that the most diagnostically informative differences between frontier models emerge not in factual recall or task completion, but in how each model handles the transition from description to commitment, and what each model independently chooses to flag as a blocking concern when none is required to do so.