Hey Product Hunt! I built ReasonPatch because AI tutors often answer before a learner has had the chance to repair their own reasoning. The interaction is deliberately small: bring an attempt, preserve what works, locate the first consequential break, ask one Socratic question, revise the exact step, then apply the idea to a fresh problem. It currently covers formal logic, algebra, Python edge cases, and causal reasoning. Under the hood, protected live mode uses GPT-5.6 Sol as planner and synthesizer, with three parallel Luna probes testing counterexamples, hidden assumptions, and rubric evidence. A failed Luna role falls back to Sol and is disclosed in the trace. The public launch uses deterministic guided fixtures with zero model calls. That is intentional: anyone can inspect the learning loop without an API key, and the demo never pretends a recorded trace is live. It also makes no claims about grades, mastery, or learning outcomes before an educator study exists. I’d especially value feedback from educators and learners: does one focused question create enough productive friction without becoming frustrating?

ReasonPatch

AI office hours that find the first break—not the answer

AI office hours that find the first break—not the answer