Proof of AI

Proof of AI (PoAI) is how Hanzo makes AI computation verifiable instead of asking you to trust a black box. It underpins two things:

• Verifiable inference — a result anyone can re-check.
• An open-source payout — a commitment that 25% of attributable cloud revenue flows back to the developers whose code the cloud runs.

This section documents what works today and what is on the roadmap. We label both, because verifiability is the entire point.

What PoAI proves, and how

Floating-point AI is non-deterministic — the same model and input can yield slightly different numbers on different hardware. You cannot reach consensus on a number that changes every run. PoAI removes the non-determinism:

• Integer-only inference. Models run in int8 with int32 accumulation, arithmetic right-shift scaling, saturating outputs, and frozen lookup tables instead of exp()/rsqrt(). The forward pass is bit-for-bit identical across CPU, Metal, CUDA, and AMD.
• Verification by recomputation. The integer transformer is a registered EVM precompile at 0x0300...0003. Every validator re-runs the identical computation during block execution; a mismatched output is rejected. No committee, no trusted hardware, no proof to forge.

Live today

• A deterministic int8 transformer (gemm_i8 keystone) implemented byte-identically in Go, Rust, and C++, with GPU kernels (Metal/CUDA/HIP) that pass the same known-answer tests.
• A registered on-chain inference precompile at 0x0300...0003, shared by the Lux, Zoo, and Hanzo EVMs.
• Cross-language golden-token tests gating CI.

Roadmap

• Production-size models. The precompile runs a small reference model today; a Qwen3-0.6B / zen-nano port is in progress. The model line is small, fully-verifiable models plus sampled verification of large ones.
• Cryptographic proofs. zkML proofs and TEE confidential execution are designed and stubbed for a later release. Today's guarantee is determinism you can re-run.
• Hardware attestation for large models. AMD SEV-SNP attestation is implemented end-to-end; Intel TDX and NVIDIA GPU confidential compute are scaffolded. See Node Operator.

A note on consensus

Hanzo and Lux chains can run Quasar, a post-quantum threshold-certificate consensus. Quasar finalizes the inference precompile's output the same way it finalizes any chain state — it does not itself inspect model output. "Verifiable inference" comes from the deterministic precompile, not from consensus.

In this section

• Node Operator — run Hanzo Node, receive work, get paid, and the trust/attestation requirements.
• Open-Source Payouts — how the cloud attributes usage to your dependencies and pays their authors.