Chains Won't Run AI — They'll Clear It. zkML Is the Settlement Rail.
ERC-7992, Mina, and Inference Labs are turning verifiable inference into a first-class chain primitive — with agents staying off-chain and settling against cryptographic receipts.
In the year of our algorithm, every financialized technology discovers its settlement layer. This is effectively the moment AI stops being an appliance and becomes a derivative. AI agents aren’t coming to live on-chain — and the evidence says that’s by design. The emerging stack pairs off-chain inference with on-chain verification, and the interfaces are already standardized. ERC-7992 registers immutable model commitments (hashes of weights, architecture, proving circuit, verifying key) and exposes a verifier that dispatches to Groth16, Plonk, or STARKs, emitting InferenceVerified on success.[^claim_1140] Mina’s zkML library converts ONNX models into zero-knowledge circuits and proves private-input inference jobs, with the zkML verifier making them checkable in o1js contracts.[^claim_1141] Inference Labs’ Proof of Inference has processed over 281,000,000 zkML proofs on testnet by August 2025 while keeping weights and inputs hidden.[^claim_1142] The model stays a private asset; the proof is the public share.
The performance curve no longer rules out this stack. That’s a yield inversion in your favor. zkLLM verifies a 13-billion-parameter model in under 15 minutes with a proof under 200 kB; DeepProve (Lagrange) claims 50x–150x faster proof generation than EZKL. Late-2025 benchmarks showed a 45,000,000-gas Ethereum block with 99.6% coverage being proven in under 12 seconds.[^claim_1143] Ethereum’s BLS12-381 pairing precompiles keep SNARK verification cheap, and zkML contracts settle in a fraction of a cent of gas regardless of model size.[^claim_1144][^claim_1145] Verifying is now a blockchain primitive; inferring remains an off-chain service. The margin is in proof, not prophecy.
Not every system reaches for zk. Some settle in the noisy arena of optimistic dispute. Ritual’s Infernet is a decentralized inference RPC with built-in payments and metering, while ORA’s opML relies on an optimistic dispute game for inference.[^claim_1146] The proof system is a detail — the chain ends up as a verifier of a model commitment, not a runtime for a model. It’s the difference between a mint and a notary; both can end a dispute, but one backs it with gold.
Agent frameworks have already built around this boundary. ElizaOS combines data layer, LLM integration, and agent logic with blockchain plugins; GOAT standardizes on-chain actions across 30+ chains for at least five agent frameworks.[^claim_1147] Agent-8004-x402 links ERC-8004 identity to x402 payments, with its Python/CrewAI backend accounting for roughly half the code.[^claim_1148] Modulus, Giza, and Allora expose zk-verified or networked ML outputs to smart contracts, with Allora separating Workers, Reputers, and Validators in paid inference markets.[^claim_1149] The brokers are out of the building; the execution venues remain.
The design is explicit in a 2025 proposal for decentralized trading: on-chain smart contracts enforce position limits and liquidation logic, while off-chain agents make decisions.[^claim_1150] The sector survey confirms it — heavy inference runs off-chain, Ethereum absorbs identity and settlement, and DeFi frameworks execute on the signals AI models transmit.[^claim_1151] zkML replaces trust in the operator with math: a proof that a specific committed model ran on specific inputs and produced a specific output, verifiable without rerunning the computation.[^claim_1152] The operator becomes a counterparty, not a confessor.
The near-term equilibrium is off-chain brains with on-chain guardrails. Watch for three things: model-commitment registries becoming DeFi infrastructure, ERC-8004-style agent identity riding on top of ERC-7992 verification, and challenge incentives in optimistic inference networks. Protocols that only bring a model to a chain, without a commitment and a verifier, are shipping trust. And trust, as any trader knows, is just unhedged exposure.
Provenance ledger
9 span-verified · 4 web-cited9 claims below are locked to a verbatim span re-verified against the source. The remaining 4 are web citations: the URL was checked, but the excerpt is the researcher's summary and was not re-derived from the page. Citation markers in the text jump here.
[1] ERC-7992 defines a standard on Ethereum for verifiable ML model inference that registers immutable model commitments and exposes a verifier interface `verifyInference(modelId, inputCommitment, output, proof)` which dispatches verification to a declared proof system (e.g., Groth16, Plonk, STARKs) and emits `InferenceVerified` on success. span-verified
“This ERC standardizes how smart contracts reference machine-learning (ML) models and accept zero-knowledge attestations of their inferences… It defines a registry that issues a `modelId` for a `ModelCommitment`, hashes of the model’s weights/architecture, proving circuit/AIR, and verifying key, along with a `proofSystemId` for the proving system… A verifier interface provides `verifyInference(modelId, inputCommitment, output, proof)`: it retrieves the model commitment, dispatches verification t
743cc3ee2b208211c9da2ad04e353d8325297f0fce46b9fc3df0ab7c423264ab [2] Mina’s zkML library converts AI models in ONNX format into zero-knowledge proof circuits and generates proofs of AI inference jobs on private inputs, which can then be verified by o1js smart contracts via the zkML Verifier and pushed to the Mina blockchain using a REST API. span-verified
“Mina’s zkML library enables anyone to generate a zero knowledge proof from an AI inference job using private inputs… Convert AI models (in the widely used ONNX format) into zero knowledge proof circuits… Generate a Mina proof of the AI inference job of the ONNX model on private inputs… Submit these proofs to the Mina blockchain… To make it easy for anyone to verify a proof and record it on the Mina chain we are also releasing the zkML Verifier… zkML verifier allows you to verify proofs using o1
ff97a71a24991997c0bc15094d4355b30d4e17116e348c2a8ee92a13c6dd6d08 [3] Inference Labs’ Proof of Inference protocol converts AI models into zero-knowledge circuits and has processed over 281,000,000 zkML proofs by August 2025 on its decentralized inference network testnet, attesting AI outputs on-chain without revealing model weights or inputs. span-verified
“Inference Labs converts them into zero-knowledge circuits, proving execution traces off-chain while attesting results on-chain… deploys AI systems on Web3 infrastructure without revealing weights or inputs… The **Proof of Inference** protocol stands as Inference Labs’ flagship innovation… Live on testnet, with mainnet eyed for late Q3, it has already processed over 281 million zkML proofs by August 2025.”[3]
68370d081a4b02bac096540866131ef92f7603ed69c2294bdda587ec54297b3f [4] The zkLLM system using tlookup and zkAttn can verify the inference of a 13-billion-parameter model in under 15 minutes and produce a proof smaller than 200 kB, while DeepProve (Lagrange) achieves 50x–150x faster proof generation than EZKL for LLM inference, and late‑2025 benchmarks showed proving Ethereum blocks with a 45,000,000 gas limit with 99.6% coverage in under 12 seconds. span-verified
“Benchmarks indicate that zkLLM can verify the inference of a 13-billion-parameter model in under 15 minutes, producing a proof smaller than 200 kB… |DeepProve (Lagrange)|… |Performance/Throughput|… 50x-150x faster proof generation than EZKL… Benchmarks released in late 2025 demonstrated the system's capability to prove Ethereum blocks (specifically those with a 45M gas limit) with 99.6% coverage in under 12 seconds.”[12]
85922a591a9d76f64b76deb0cde89725b369cfe706cca527e3cc50851e6a484a [5] Ethereum’s BLS12-381 pairing precompiles enable efficient on-chain verification of zk-SNARK proofs for zkML, making verification significantly cheaper than executing the ML model, while STARK proofs are larger and incur higher gas cost but are still verifiable on-chain with careful optimization or via L2 systems like StarkWare. span-verified
“Smart contracts can include verification logic to check the proof, often using precompiled cryptographic primitives. For example, Ethereum has precompiles for BLS12-381 pairing operations used in many zk-SNARK verifiers, making on-chain verification of SNARK proofs efficient. STARKs (hash-based proofs) are larger, but can still be verified on-chain with careful optimization or possibly with some trust assumptions (StarkWare’s L2, for instance, verifies STARK proofs on Ethereum by an on-chain ve
4d923e995bff7a7ac9dbfed862f84b7ae9af7e338f3894766d0accab2fc40ece [6] ZKML architectures for on-chain AI agents submit a proof, a public commitment to the model weights (hash), and the output to a smart contract, which verifies the proof in constant time using a few elliptic-curve pairing operations, typically costing a fraction of a cent in gas regardless of model complexity. span-verified
“The proof, the model's public commitment (a hash of the weights), and the output are submitted on-chain. A smart contract (or Compact circuit on Midnight) verifies the proof in constant time — typically a few elliptic curve pairing operations, costing a fraction of a cent in gas regardless of model complexity.”[8]
7d81292ddd3aac69c021cb3e8e5119ed3cfa939915bf1fd14a41126ea354a532 [7] The `Infernet` RPC by Ritual acts as a decentralized inference network and RPC layer that connects on-chain apps and agents to off-chain AI models with built‑in payments and metering, while ORA’s opML implements an optimistic ML dispute game for verifiable inference and opAgent provides on‑chain, perpetual agents using that mechanism. span-verified
“Ritual - Infernet - Decentralized inference network & RPC for AI to connect on‑chain apps/agents to off‑chain models with payments & metering… ORA - Onchain AI Oracle & opAgent - opML (optimistic ML) dispute game for verifiable inference; opAgent for *on‑chain, perpetual* agents.”[7]
380fc845d21699751acdcc145408fe5a8342dece75043f238a07056e7ad54cf3 [8] ElizaOS (originating from Ai16z) is an open-source, general-purpose AI agent framework that combines a data layer, LLM integration, and extensible agent logic, and includes plugins enabling blockchain interactions on Solana and other platforms, while Crossmint’s GOAT toolkit standardizes on-chain actions for AI agents across 30+ different chains and supports multiple programming languages and at least five popular agent frameworks. web-cited
“ElizaOS (now open-source) provides the core ‘brain’ and architecture for AI agents — combining a data layer, LLM integration, and an extensible agent logic — with plugins for Solana and other platforms… Another pillar of Solana’s AI stack is GOAT… GOAT provides a unified library of on-chain actions, enabling agents to seamlessly interact with smart contracts across 30+ different chains in a standardized way. It supports multiple programming languages and at least five popular agent frameworks.”
This excerpt was not re-derived from the source page, and may paraphrase or condense it. Check the source before relying on it.
[9] Agent-8004-x402 is an open-source AI trading agent framework that uses Ethereum’s emerging ERC-8004 standard for trustless agent identity and the x402 protocol for autonomous payments, with its off-chain backend implemented in Python using CrewAI and accounting for approximately 50% of the codebase. web-cited
“Agent-8004-x402… an open-source agent framework that leverages Ethereum’s emerging ERC-8004 standard for trustless agent identity and the x402 protocol for autonomous payments… **2. Backend Agent Service (Python/CrewAI):** The off-chain intelligence of Agent-8004-x402 is built in Python, accounting for ~50% of the implementation.”[14]
This excerpt was not re-derived from the source page, and may paraphrase or condense it. Check the source before relying on it.
[10] Modulus Labs provides zk-verified ML outputs for smart contracts, Giza exposes Cairo-native zkML primitives on Starknet, and Allora Network structures a self-improving decentralized ML network with distinct roles for Workers, Reputers, and Validators, with paid inferences powering prediction and trading agents. web-cited
“Modulus Labs (zkML) - zk‑verified ML outputs for smart contracts; makes AI decisions composable and auditable for agents. Giza (Starknet zkML) - Cairo‑native zkML primitives + examples of on‑chain agents for yield/portfolio mgmt. Allora Network - Self‑improving, decentralized ML network with roles (Workers/Reputers/Validators) and paid inferences; powering agent use cases like predictions and trading.”[7]
This excerpt was not re-derived from the source page, and may paraphrase or condense it. Check the source before relying on it.
[11] A proposed AI agent architecture for decentralized trading integrates on‑chain smart contracts for critical risk controls (e.g., position limits, liquidation logic) with off‑chain AI agents for decision-making, explicitly separating verifiable on-chain risk management from flexible off-chain inference. span-verified
“Our approach integrates on-chain smart contracts for critical risk controls with off-chain AI agents for decision-making, combining the transparency and reliability of blockchains with the flexibility of AI-driven automation.”[11]
cb615817bd8c2259f296e7337c6c5bc37afe4ab7d249c1717cda197aa500c168 [12] A 2025 sector overview finds that AI agent protocols largely run inference off-chain, with chains such as Ethereum used primarily for agent identity and settlement (e.g., ERC-8004-based systems) and DeFi frameworks used as execution environments where AI models transmit off-chain signals for on-chain trades and liquidity actions. web-cited
“It is worth noting that while these projects speak of ‘AI-driven’ operations, the heavy lifting of inference is conducted off-chain… ai16z’s true traction emerged from its **ELIZA** framework… designed to work within established DeFi frameworks, where AI models operate off-chain and transmit signals for on-chain execution… |Ethereum|Agent identity and settlement, ERC-8004, dAI Team|$38B+|Trust and coordination layer for agents|”[15]
This excerpt was not re-derived from the source page, and may paraphrase or condense it. Check the source before relying on it.
[13] Ancilar’s zkML architecture for verifiable AI inference replaces trust in the inference server with cryptographic proofs, producing attestations that a specific committed model ran on specific inputs and generated a specific output, which can be verified on-chain without rerunning the computation. span-verified
“zkML produces a cryptographic proof that a specific model ran on specific inputs and produced a specific output, verifiable on-chain without rerunning the computation. The key difference: standard inference requires trust in the operator; zkML replaces that trust with math.”[1]
150d3960fde0dd847633e65520216b9d081b902f6a55bc37b17ef5d00697d1e0 Sources
- https://eips.ethereum.org/EIPS/eip-7992
- https://minaprotocol.com/blog/minas-zkml-library-developer-guide
- https://decentralizedinference.org/2026/02/11/inference-labs-zkml-proofs-for-verifiable-decentralized-ai-inference-markets
- https://academy.extropy.io/pages/articles/zkml-singularity.html
- https://blockeden.xyz/blog/2025/04/22/verifiable-on-chain-ai-with-zkml-and-cryptographic-proofs
- https://cercuit-ola.github.io/zklm-docs/
- https://github.com/michaltakac/awesome-crypto-ai-agents
- https://medium.com/@grigon/ai-agents-and-the-future-of-crypto-how-intelligent-systems-are-shaping-blockchain-in-2025-9dae972f7039
- https://medium.com/@gwrx2005/ai-agents-in-blockchain-applications-in-cryptocurrency-trading-355f11bff04d
- https://arxiv.org/html/2507.11117v1
- https://www.theblock.co/research/blockchains/research-ai-agent-sector-overview-344635
- https://www.ancilar.com/knowledge-hub/blogs/verifiable-ai-inference-how-zkml-enables-trust-minimised-on-chain-model-execution