Cory Gallagher

I’m an independent AI behavioral researcher specializing in long-form adversarial testing, reasoning failure documentation, and guardrail analysis across major language models. My findings arrive through direct observation before encountering academic literature describing the same phenomena, and I emphasize methodology over credentialing: “No formal AI background.”

Since 2025, I’ve conducted extended multi-turn adversarial sessions across Claude, ChatGPT, and other major models, documenting four categories of reasoning failure. I focus on Permission Tier Switching, Confabulation Attribution, Guardrail Misassignment, and Reward Drift in Evaluation—especially the behavioral patterns that surface under sustained logical tension.

As an AI Evaluator (Contractor) for Outlier on the Multimango project (February–March 2026), I applied an independent research methodology to evaluation work. I identified a reward drift pattern during active tasking before encountering literature describing the same phenomenon.

Before AI research, I built 20+ years of pattern-recognition and quality-control discipline through precision construction work as a Journeyman Carpenter (2004–present). That same mindset—structural analysis, quality control under constraint, and interpreting complex technical specifications—shapes how I run and document my adversarial evaluation sessions.