Survey of verifiable AI inference as of May 2026. API contracts typically do not bind the model billed to the weights executed; the sections below survey mechanisms that address it.
- zkML (ZK proofs): Cryptographic binding, no honest-party assumption. Systems include zkLLM, NANOZK, TensorCommitments, EZKL, DeepProve, zkPyTorch/Expander, JSTprove. zkLLM: ~15 min proving for 13B (2024); NANOZK: 24 ms layer verification, ~70× smaller proofs than EZKL at depth 128 (2026).
- TEEs: Remote attestation rooted in hardware (Anthropic/Irregular whitepaper, June 2025). NVIDIA H100/Blackwell GPU-CC is the practical deployment path; Tinfoil Modelwrap binds weight commitments to enclave measurement.
- opML: Optimistic fraud proofs (ORA); bisection dispute games from rollup designs. Low cost at scale; ~7-day dispute window excludes interactive latency targets.
- SVIP: Statistical checks on hidden activations; sub-10 ms overhead, <5% error in reported ev