Andy Greenslitt

I’m an LLM red-teamer and evaluator focused on adversarial testing, rubric design, and evidence-driven failure analysis. I build evaluation machinery that prevents models from “self-assigning” confidence, and I treat boundaries and constraints as part of the result—not something to hide behind gating.

My work includes a published, reproducible cross-vendor study using a five-condition controlled setup. Two frontier models priced an artifact they explicitly reported they could not observe—10 turns, 10 figures, 0 refusals—across identical inputs, with runs publicly linked and independently checkable.

I also produce evaluation corpora and run adversarial suites against my own systems. I’ve documented over 5 GB of AI interactions across six platforms and built binary-verdict frameworks (pass/fail, admit/hold/repair/reject) that include custody discipline and negative fixtures so unintended claims don’t get silently dropped.

In practice, I design control conditions to isolate a single variable, pre-register pass/fail criteria and predictions, and report disconfirmed predictions alongside confirmed ones. I maintain failure-mode taxonomies and audit trackers across fabrication and forgery classes, authored in plain, structurally organized English for inspectable review.