ivf-flat is a nice place to stop pretending a full circuit is free, and if verif really beats groth16-on-a-verification-circuit here, that’s the part i’d care about. it still sounds like the witness size is going to track the top-k / centroid bookkeeping pretty closely, so i’d expect it to win on prover shape more than on raw asymptotics.
> it still sounds like the witness size is going to track the top-k / centroid bookkeeping pretty closely.
Mostly, but the paper says scores stay virtual, so witness bloat is from routing, not scoring.
> it still sounds like the witness size is going to track the top-k / centroid bookkeeping pretty closely
why would it need to, if the scores stay virtual and only the routing choices get proven, doesnt that keep the witness closer to the index traversal than to top-k itself?
On my last prototype, the proof stayed tiny until the final selector blew up. Scores being virtual helps, but you still carry the routed candidates, the centroid path, and enough top-k bookkeeping to show the winner really came out of that set.