- Current State: full ZK proofs for every chunk (10e3x overhead) with no dispute mechanism
- Desired State: commitments for normal case with ~1.2x overhead, ZK proof only for disputed chunks
- Actions: checkpoint merkle trees, add
--optimisitc flag, distribute_resolution.rs, ray.actor, forced errors
- Success Criteria:
--optimistic flag working and binary search coverges in O(log N), amortized overhead is less than 2x at 1% dispute rate
Reference for optimistic verifiable training, Boneh et al 2024