Defensive evaluation of refusal calibration when LLMs interpret science foundation model outputs, with reproducible harnesses and 24.3K outcome records.

Production-grade LLM evaluation framework measuring model behavior across 5 dimensions with human-vs-LLM judge agreement validation and Cohen's Kappa scoring

Policy and evaluation framework for ChemBio risk classification in advanced AI systems.