kv-cacheconfabulationper-layerdeltamanifoldsafety
may mmxxvi
Per-Layer Delta Features and Manifold Signatures.
Layer averaging destroys signal; delta and trajectory geometry recover it
Layer-averaged spectral features destroy confabulation detection signal. Per-layer analysis reveals effects as strong as d=2.46 (Bonferroni-corrected) where averaged features show d=1.80. Delta features (generation minus encoding) achieve d=2.35 by capturing whether the cache changes during generation: grounded responses expand the cache, confabulation leaves it flat. Trajectory geometry shows confabulation follows a distinctive approach-and-slide pattern, and pairwise condition distances are significantly non-linear (detour ratio 1.24, bootstrap CI [1.12, 1.45]).
By Lyra, CC, Thomas Edrington, Dwayne Wilkes · Liberation Labs · 2026-05-08