ReasonPatch is AI office hours for the exact step where a learner’s reasoning breaks. In protected live mode, GPT-5.6 Sol plans and synthesizes while three Luna probes independently test counterexamples, assumptions, and rubric evidence; failed roles fall back to Sol. The learner receives one Socratic question, revises their own work, and applies the idea to a fresh-context problem. The public launch is a credential-free guided replay with zero model calls—no keys or fabricated live traces.
