White paper WP-016
Verifiable ML Inference: ZK Proofs as Output Attestation
Show an inference happened correctly when you cannot show the model.
At a glance
- Paper
- WP-016
- Topic
- Verifiable ML Inference: ZK Proofs as Output Attestation
- Format
- PDF + web summary
- Signatures
- ML-DSA-65 post-quantum (NIST FIPS 204)
- Sandbox
- Reproducible at affix-io.com/sandbox
- Company
- AffixIO, Wales, UK
Verifiable ML Inference: ZK Proofs as Output Attestation is an AffixIO technical paper. Show an inference happened correctly when you cannot show the model.
Third parties need to trust model outputs in credit, insurance, and public-sector AI. ZKML lets you prove a specific inference satisfied policy constraints without publishing weights or user data. We survey what is production-ready in 2026 and what still belongs in the lab.
Summary
Third parties need to trust model outputs in credit, insurance, and public-sector AI. ZKML lets you prove a specific inference satisfied policy constraints without publishing weights or user data. We survey what is production-ready in 2026 and what still belongs in the lab.
Download the full PDF for technical detail, diagrams, and reproduction steps. Public sandbox: affix-io.com/sandbox.
Related reading
Frequently asked questions
What is verifiable ML inference?
A cryptographic proof that a named model version produced a stated output on committed inputs, verified without exposing those inputs.
Is ZKML fast enough for production?
Small circuits verify in milliseconds; proving still costs hundreds of milliseconds to seconds depending on model size and hardware.
Which tools does AffixIO use?
constraint language and proving backend for governance circuits, with EZKL and related stacks for ML-specific workloads where circuit size allows.