I'm an independent researcher based in California working on AI safety evaluation, with a focus on whether safety-relevant model behaviour holds across sustained interaction rather than single turns.

I developed TRACE-FV, a preregistered black-box protocol testing whether verified corrections to AI self-claims remain operative under frame variance (DOI: 10.17605/OSF.IO/6U3QX). Protocol, analysis plan, and implementation are public. I'm particularly interested in falsifiable evaluation methods, reproducible experiments, and collaborations that turn underexplored model-behaviour questions into rigorous empirical work.