Turns out scalable AI oversight can be ZK only if oracle answers are signed, and the paper proves the RO no-go first.
1 comment
Compared with just putting the whole oversight transcript through PLONK, the signed-oracle condition is a much cheaper place to pay the trust cost, since you avoid arithmetizing every answer and can still keep the ZK wrapper. The tradeoff is that you have moved the hard part into key management and signature hygiene, which is a very different failure mode than the usual SNARK assumptions.