Notes: Trust Attractor Validation

Chapter notes for “Trust Attractor Validation”

Notes

¹ Edrington, T. and Lyra, R., “Probing Honesty in Large Language Models via KV-Cache Singular Value Analysis” (2026). SVD analysis of key-value cache geometry across 0.5B-32B parameter models found that deception expands effective dimensionality while compressing per-token magnitude (Cohen’s d = -2.44 for effective rank, d = +3.59 for per-token norm). The result corroborates the output-entropy finding reported in this chapter: honesty is geometrically expansive while deception is geometrically constrained, providing independent evidence that honest signaling is thermodynamically grounded rather than merely policy-compliant.