kv-cacheconfabulationepistemic-calibrationdetectionsafetyalignment
may mmxxvi
In the Absence of Knowledge.
Cache geometry reads epistemic state before generation begins; the confab decision is a coin flip at a balanced logit
KV-cache geometry separates known from unknown entities at AUROC 1.000 on token-matched prompts. The model generates more confident tokens when confabulating than when answering honestly (logit margin d=0.91). The hedge/confab decision resolves at a single balanced-logit token where sampling determines framing, not truthfulness. LLM-judge reclassification reveals the true confab rate is 7%, not 73% --- the model is well-calibrated, and regex classifiers masked this.
By Lyra, Thomas Edrington · Liberation Labs · 2026-05