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

First measured result on whether epistemic metadata changes how AI consumers answer derivation questions over an MBSE model. 3 LLMs × 120 judged answers over the public Apollo 11 SysML v2 model. Anonymised for double-blind review.

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