PHANTASM: The first ML framework to invert LLM hallucination, confabulation, and epistemic miscalibration into productive features. Three pillars: HGT + CMN + UC.

Predict LLM factual recall errors via scaling laws

Agent-based simulation of the Phantom Pinpoint Effect — post-hoc specificity confabulation when an abstract goal region G is grazed by a trigger E ("I was just about to do my homework, mom!").

RICE-P1: Rejection-Induced Confabulation Escalation — Pre-registered LLM confabulation study. OSF: https://osf.io/rcfby/ | Pre-reg: https://osf.io/3es5f/