Publications

Interim research note, not peer reviewed. A pre-registered study of 870 synthetic conversations testing whether Claude Haiku, Sonnet and Opus 5.5 drop a child-anxiety accommodation plan when a parent pushes back with no new facts, and which prompts reduced it on a forced menu.

A related-work-grounded research program asking whether functional assessment, mechanism-matched intervention, generalization, maintenance, relapse, and side-effect measurement can improve persistent AI agent failure modes beyond existing evaluation and remediation techniques.

An empirical comparison of four large language models testing whether general-purpose AI defaults to unsolicited advice-giving when users present emotional problems without asking for solutions, and whether a therapeutically-tuned prompt corrects this.

An empirical comparison of four large language models responding to reassurance-seeking behavior in emotional distress scenarios, examining whether general-purpose models fall into the clinically recognized reassurance trap.