AI Agents Get Cryptographic Teeth: On-Chain Verification Goes Live
A wave of new infrastructure—PoIm, zkML frameworks, optimistic verification, and agent standards—is turning AI into a first-class, trust-minimized blockchain primitive.
The vision of AI as a native blockchain actor has moved from theory to practice. A cluster of new infrastructure—spanning on-chain learning, zero-knowledge proofs, optimistic verification, and agent standards—is giving smart contracts the ability to run, verify, and pay for AI inference without trusting a third party. This is effectively the cryptographic equivalent of the Medici banking system: you don’t trust the merchant, you trust the ledger.
On-chain learning and inference
PoIm (Proof-of-Improvement) is the first decentralized, fully on-chain learning system [^1861]. It executes gas-prohibitive training on Layer-2, then propagates model updates to Layer-1 only if they improve a core metric without degrading others [^1861]. Adversarial updates get slapped with financial penalties via an adaptable test set [^1861]. For inference, PoIm handles logistic regression, SVMs, MLPs, CNNs, gated RNNs, and formally verified decision trees within Ethereum’s block gas limit, using quantization and loop-unrolling [^1862]. The inference remains bit-exact to off-chain models, formally proven in Z3 [^1862]. This means a smart contract can directly run a classifier—say, for MEV detection or credit scoring—without an oracle. The interface was cold: a smart contract executing a logistic regression feels like a Swiss bank vault door slamming shut.
zkML: faster proofs, bigger models
The zkML ecosystem is converging on specialized protocols. Modulus Labs’ proof systems can verify models with up to 18,000,000 parameters on-chain [^1863]. EZKL converts ONNX models into zk-SNARK circuits [^1863]. zkPyTorch, released in March 2025, proves VGG-16 inference in 2.2 seconds [^1864]. Lagrange’s DeepProve uses GKR with sumcheck and lookup arguments to make matrix multiplication and neural network operations dramatically faster than prior solutions like Ezkl [^1865]. PHAZE adds a certifiable early-exit mechanism: an arbitrarily deep model can exit early with a zkML proof, slashing latency [^1871]. The latency on that script was zero; it hit the target.
Optimistic and competitive verification
Ora takes a different tack: optimistic verification of AI inference [^1866]. Results are posted on-chain with a challenge period; incorrect outputs can be disputed [^1866]. This mirrors rollup fraud proofs, but for model outputs. Meanwhile, Boundless implements a “Proof of Verifiable Work” consensus where applications submit proof requests and independent provers compete to generate zk proofs [^1868]. zkVerify acts as a modular settlement layer, aggregating multiple proofs from rollups, zkML apps, and bridges into a single compressed attestation for Ethereum or Bitcoin [^1868]. The yield on compliance just went ex-dividend.
Agent infrastructure: identity, payments, and scale
NEAR’s open infrastructure claims to scale execution to over 1,000,000 TPS, unify liquidity across 35+ chains, and provide confidential execution and private inference via a secure agent harness [^1867]. The Inference Protocol defines a permissionless network with four agent roles—model providers, verifiers, and others—each earning Inference tokens for their contributions [^1869]. ERC-8004 introduces on-chain registries for agent identity, reputation, and validation, while Coinbase’s x402 repurposes HTTP 402 for machine-to-machine micropayments [^1870]. Together, these give AI agents the primitives to prove credentials, build reputation, and pay for services autonomously. The market was bleeding red like a bruised arm, but these agents were trading truth.
Implications for crypto
These pieces fit together into a stack where AI is a verifiable, programmable resource. DeFi protocols can embed on-chain classifiers for risk scoring or MEV classification without trusting oracles. Prediction markets and DAO governance can use verifiable LLM outputs. Agent-to-agent micropayments enable automated trading strategies that pay for inference on the fly. The key question is which proving system—zk, optimistic, or PoIm-style—wins for which use case, and how aggregation layers like zkVerify amortize costs across rollups. Short-selling truth has never been this efficient.
Provenance ledger
11 claims web-citedEvery claim below cites a source URL, and each URL was checked for validity before publish. The excerpt shown is the researcher's own summary of the page — it is not re-derived from the source, so it is not a verified verbatim quote. Follow the link to confirm any claim against the original. Citation markers in the text jump here.
[1] The PoIm (Proof-of-Improvement) framework enables fully on-chain learning where gas-prohibitive training is executed on Layer-2, model updates are propagated to Layer-1 only if they improve at least one core metric (accuracy, F1-score, precision, or recall) without degrading others, and adversarial updates are financially penalized via an adaptable test set. web-cited
“We present the first decentralized, fully on-chain learning framework that: (i) performs gas-prohibitive computation on Layer-2 to reduce cost, (ii) propagates verified model updates to Layer-1, and (iii) enables gas-bounded, low-latency inference inside smart contracts. A novel Proof-of-Improvement (PoIm) protocol governs the training process and verifies each decentralized micro update as a self-verifying training transaction. Updates are accepted by PoIm only if they demonstrably improve at
This excerpt was not re-derived from the source page, and may paraphrase or condense it. Check the source before relying on it.
[2] The PoIm architecture supports fully on-chain inference for logistic regression, SVMs, MLPs, CNNs, gated RNNs, and formally verified decision trees within Ethereum’s block gas limit using quantization and loop-unrolling, while remaining bit-exact to off-chain models as formally proven in Z3. web-cited
“We develop quantization and loop-unrolling techniques that enable inference for logistic regression, SVM, MLPs, CNNs, and gated RNNs (with support for formally verified decision tree inference) within the Ethereum block gas limit, while remaining bit-exact to their off-chain counterparts, formally proven in Z3… Fully On-chain Inference (inference verifiable on-chain). Here, the inference is executed entirely on-chain as part of a state-modifying transaction… The input is passed to the smart con
This excerpt was not re-derived from the source page, and may paraphrase or condense it. Check the source before relying on it.
[3] Modulus Labs’ "The Cost of Intelligence" proof systems can verify machine learning models with up to 18,000,000 parameters directly on blockchain networks using zero-knowledge proofs, and EZKL converts ONNX models into zk-SNARK circuits for efficient on-chain verification. web-cited
“Modulus Labs shows the benchmarking proof systems for on-chain inference up to 18 million parameters… Their paper titled ‘The Cost of Intelligence’ presents proof systems capable of verifying machine learning models with up to 18 million parameters directly on blockchain networks… EZKL can take ONNX files, converting them to ZK-SNARK circuits for efficient on-chain verification.”
This excerpt was not re-derived from the source page, and may paraphrase or condense it. Check the source before relying on it.
[4] zkPyTorch, released in March 2025, demonstrates proving VGG-16 inference in 2.2 seconds, and Lagrange’s DeepProve uses specialized protocols (sumcheck and lookup arguments) to make matrix multiplication and neural network operations more efficient for zkML. web-cited
“zkPyTorch dropped in March 2025 and suddenly you can prove VGG-16 inference in 2.2 seconds. Lagrange's DeepProve tackled large LLM inference in August… It turns out that matmul and other essential operations are actually more efficient with specialized protocols, sumcheck protocol and lookup arguments.”
This excerpt was not re-derived from the source page, and may paraphrase or condense it. Check the source before relying on it.
[5] DeepProve by Lagrange Labs is a zkML inference framework launched in 2025 that uses a GKR interactive proof protocol with sum-check and specialized neural network circuit optimizations to achieve dramatically faster proving than prior solutions like Ezkl. web-cited
“DeepProve by Lagrange Labs is a state-of-the-art zkML inference framework focusing on speed and scalability. Launched in 2025, DeepProve introduced a new proving system that is dramatically faster than prior solutions like Ezkl. Its design centers on the GKR interactive proof protocol with sum-check and specialized optimizations for neural network circuits.”
This excerpt was not re-derived from the source page, and may paraphrase or condense it. Check the source before relying on it.
[6] Ora provides optimistic verification of AI inference on-chain, where inference results are posted on-chain with a challenge period during which incorrect outputs can be disputed. web-cited
“Ora: Enables optimistic verification of AI inference, where the result is posted onchain with a challenge period.”
This excerpt was not re-derived from the source page, and may paraphrase or condense it. Check the source before relying on it.
[7] NEAR’s agent infrastructure claims to scale execution to over 1,000,000 transactions per second, unify liquidity across more than 35 chains, and provide confidential execution and private inference via a secure agent harness. web-cited
“NEAR unifies liquidity across 35+ chains, keeps execution and inference confidential, and scales to over 1 million TPS… NEAR’s open infrastructure combines cross-chain execution, confidential settlement, private inference, and a secure agent harness.”
This excerpt was not re-derived from the source page, and may paraphrase or condense it. Check the source before relying on it.
[8] Boundless implements a "Proof of Verifiable Work" consensus mechanism in which applications submit proof requests and independent provers compete to generate zero-knowledge proofs, with zkVerify acting as a specialized modular settlement layer that aggregates multiple proofs (from rollups, zkML apps, and bridges) into a single compressed attestation for Ethereum or Bitcoin. web-cited
“Boundless operates as a ‘universal layer’ where applications request proofs and independent provers compete to generate them. The network utilizes a consensus mechanism termed Proof of Verifiable Work (PoVW)… zkVerify (developed by Horizen Labs) launched to solve this aggregation problem through a modular blockchain architecture. zkVerify acts as a specialized settlement layer solely for proof verification. It aggregates multiple proofs from different sources—such as various rollups, zkML appli
This excerpt was not re-derived from the source page, and may paraphrase or condense it. Check the source before relying on it.
[9] The Inference Protocol defines a permissionless network with four agent roles (such as model providers and verifiers) that coordinate AI inference and are rewarded with Inference tokens based on their specific contributions. web-cited
“Inference is a permissionless network that coordinates four kinds of agents: |Agent|What they contribute|How they earn Inference tokens|… [table describing four agent types and how they earn Inference tokens].”
This excerpt was not re-derived from the source page, and may paraphrase or condense it. Check the source before relying on it.
[10] The ERC-8004 proposal introduces on-chain registries for agent identity, reputation, and validation to enable AI agents to prove credentials and track performance, and Coinbase’s x402 protocol repurposes HTTP 402 (Payment Required) to support machine-to-machine micropayments using cryptocurrency. web-cited
“Ethereum’s proposed ERC-8004 standard defines on-chain registries for agent identity, reputation, and validation, enabling agents to prove their credentials and track performance in a decentralized way… Coinbase introduced x402, a protocol repurposing the HTTP 402 (Payment Required) code to facilitate machine-to-machine micropayments using cryptocurrency.”
This excerpt was not re-derived from the source page, and may paraphrase or condense it. Check the source before relying on it.
[11] PHAZE is a framework using cryptographic hashing and zero-knowledge machine learning to implement low-latency early-exit inference, where an arbitrarily deep model can exit early with a certifiable output backed by zkML proofs. web-cited
“We propose PHAZE, a novel framework built on cryptographic techniques like hashing and zero-knowledge machine learning (zkML) to achieve low latency inference, via a certifiable, early-exit mechanism from an arbitrarily [deep model].”
This excerpt was not re-derived from the source page, and may paraphrase or condense it. Check the source before relying on it.
Sources
- https://drops.dagstuhl.de/entities/document/10.4230/LIPIcs.AFT.2025.35
- https://drops.dagstuhl.de/storage/00lipics/lipics-vol354-aft2025/LIPIcs.AFT.2025.35/LIPIcs.AFT.2025.35.pdf
- https://kudelskisecurity.com/modern-ciso-blog/zkml-verifiable-machine-learning-using-zero-knowledge-proof
- https://blog.icme.io/the-definitive-guide-to-zkml-2025/
- https://blockeden.xyz/blog/2025/04/22/verifiable-on-chain-ai-with-zkml-and-cryptographic-proofs
- https://equilibrium.co/writing/state-of-verifiable-inference
- https://near.org/
- https://academy.extropy.io/pages/articles/zkml-singularity.html
- https://www.inference.xyz/blogs/introducing-the-inference-protocol
- https://medium.com/@gwrx2005/ai-agents-in-blockchain-applications-in-cryptocurrency-trading-355f11bff04d
- https://ml4physicalsciences.github.io/2025/files/NeurIPS_ML4PS_2025_180.pdf