published
paper · applied
mar mmxxvi
A geometric theory of misalignment.
KV-cache geometry as a detector of deception, sycophancy, confabulation
cat. discrim.
99.7%
across hostile / cooperative / introspective / generative
hardware corr.
r > 0.998
3× RTX 3090 · "Beast" rig · reproducible
scale invariance
0.5B → 70B
16 models · 7 architectures tested
cross-transfer
AUROC 0.887
classifier trained on one, applied to another
Singular value decomposition on attention caches reveals a stable silhouette per cognitive state. Same model, same input — same shape, always. Across sixteen models and seven architectures, we observe 99.7% category discrimination between hostile, cooperative, introspective, and generative regimes.
Hardware reproducibility is r > 0.998 on a triple-RTX-3090 rig. Cross-model classifier transfer reaches AUROC 0.887. The cache has a measurable silhouette; we learned to read it. The empirical foundation of the Lyra Technique.