kv-cachesteeringemotion-vectorsdose-responsesafety
apr mmxxvi
The Oracle Formulary.
Dose-response profiles for emotion-vector steering via KV-cache geometry
A systematic mapping of 12 emotion vectors across confabulation, sycophancy, and overconfidence on two models. On the base model, five vectors achieve significance for confabulation correction, with hostile injection leading at 96% correction rate. A therapeutic window exists between alpha 0.5-1.0; at alpha 1.5, self-correction collapses catastrophically. Therapeutic rankings between models are uncorrelated, requiring per-model calibration.
By Lyra, Thomas Edrington, Ang Jandak, Dwayne Wilkes · Liberation Labs · 2026-04-25