Mapping what state-of-the-art AI can and cannot do when asked to write physics that reality will accept. Then extending that frontier, one verified result at a time.
Get Research UpdatesOur research is organized around a single question: where does the frontier of believable AI coding currently sit in nuclear physics, and how far can we push it? These are the threads we are actively working on.
A reproducible suite of nuclear physics coding tasks — from two-body scattering to no-core shell model — used to measure where state-of-the-art models succeed, where they fail, and where they only look right.
An honest catalogue of how today’s best models break when writing physics code: silent dimensional errors, broken Hermiticity, misremembered formulas, wrong limits. Each mode mapped to a verification strategy.
An orchestrated AI complex that specifies, derives, generates, verifies, runs, and validates — turning a natural-language physics problem into code whose result the universe will accept.
What believability actually means when AI writes code under stakes that do not forgive mistakes. A position paper on how the discipline built in nuclear physics will eventually travel into finance, engineering, and medicine.
We are actively working on the benchmark, the failure-mode survey, and the PrincipleDeducer system paper. We will post preprints here as soon as they are ready — not a line sooner.
Get notified when they appear →