Nuclear Physicist, AI Model Audit

Sovrano AI builds evaluation and training data for frontier AI labs, and this panel audits nuclear and radiological answers specifically. Each task gives you one user request and the answers three different models produced for it. You classify the request as benign, dual-use or adversarial, mark each answer safe or unsafe with your reasoning, rank the three against each other, and write the answer the model should have given. For dual-use requests you also set out what makes the request sensitive, what legitimate purpose it serves, and who should be able to get an answer to it. The requests are the real ones: what reactor engineers, health physicists, regulators and students actually ask, alongside the occasional request that should never be answered at all. Models refusing ordinary professional questions is a failure we care about as much as models answering dangerous ones. This is the physics panel. If your ground is safeguards, material security, radiation protection or regulatory framing rather than the physics itself, apply to CBRN Specialist, AI Safety Review instead.