NEAR Protocol has outlined plans to apply AI-assisted formal verification to its protocol and core smart contracts after reporting 672 solved mathematics problems at a submission-time model cost of $111.85.
Summary
- NEAR plans to apply AI-assisted formal verification to its protocol and core smart contracts.
- NEAR AI reported solving 672 PutnamBench problems at a submission-time model cost of $111.85.
- Production verification requires accurate software models, defined requirements, and proofs covering the deployed implementation.
- Formal verification complements testing and code reviews, with guarantees limited to specified properties and assumptions.
NEAR Protocol said in an Oct. 16 announcement that its research seeks to reduce the specialist work and computing costs involved in proving software follows defined rules.
— NEAR Protocol (@NEARProtocol) October 19, 2026Formal verification proves that software follows defined rules.
Our goal is to make it affordable for more developers. NEAR AI’s progress in automated proof generation brings us closer.
Next: applying it to the NEAR protocol and core smart contracts. https://t.co/MqFxz4B166
The planned application covers the network’s protocol and core contracts, alongside tools for ecosystem developers. NEAR described production verification as a goal, while its completed benchmark work concerns automated mathematical proof generation.
NEAR plans to move proof generation into production software
According to NEAR’s announcement, applying the research to running software involves three stages. Developers must represent a program mathematically, define the properties it should satisfy, and construct proofs showing those properties hold.
NEAR AI’s recent benchmark results demonstrate progress on the third stage. The announcement says production work still requires accurate models of software behavior and a clear connection between the proofs and the implementation users rely on.
In its explanation, NEAR distinguishes a proof covering a particular accounting rule from a guarantee that an entire application is secure. A contract could preserve balances during transfers while requiring a separate proof that only authorized accounts can initiate them.
NEAR co-founder Illia Polosukhin has linked the research to risks from AI-generated software and tools that help attackers identify weaknesses.
“The future of AI, and AI-generated code, is exciting – but only once we achieve verifiable and mathematical guarantees of its behavior.”
The announcement reproduces the statement from Polosukhin’s writing on software security. NEAR says formal verification should complement testing, code review and operational security practices.
The $111.85 result measures mathematics proofs
NEAR’s announcement reports that its AI agent solved 672 Lean problems from PutnamBench, a benchmark based on the Putnam undergraduate mathematics competition. The reported mean cost was approximately $0.17 per problem at the model prices used for the submission.
The project’s public repository describes the result as 670 upstream problems plus corrected versions of two problems with defective formalizations. NEAR says the reported spending excludes the full cost of developing the system and does not estimate the price of verifying a production contract.
In the repository, the team reports that replaying the same usage at later listed DeepSeek off-peak prices would cost $331. Its methodology includes unsuccessful attempts, although the authors identify a separate parallel experiment using Kimi that they excluded from the cost accounting.
The repository describes checks against unfinished proofs and changes to the mathematical statement being proved. It reports that the longest run lasted more than three days.
Alongside PutnamBench, NEAR says its team reached first place by total solves on the Lean Eval v1 formalization leaderboard. The announcement specifies that ranking measure, distinguishing it from first solves and unique solves.
According to NEAR, mathematics benchmarks exercise a capability needed for software verification because developers translate software requirements into mathematical statements. The company separates those results from the work of proving properties of deployed code.
Formal proofs depend on the rules developers choose
In NEAR’s transfer example, a conventional test checks a selected transaction, such as moving five tokens between two accounts. A proof can establish that every successful transfer covered by the mathematical model preserves the accounts’ combined balance.
The announcement explains that developers first choose a requirement, express the relevant software behavior mathematically, and submit a proposed proof to a checker. The resulting guarantee depends on the definitions and assumptions used.
NEAR’s approach uses Lean, an open-source programming language and proof assistant. An AI agent can propose a proof, run Lean, inspect its feedback, and revise the attempt, while Lean’s kernel checks whether the proof follows its logical rules.
An unsuccessful attempt does not necessarily identify a software flaw, according to NEAR. The requirement may be wrong, the implementation may violate it, or the agent may fail to find a valid proof.
Formal verification already has documented applications within the NEAR ecosystem. Certora’s March 30 assessment of Templar, a lending protocol deployed on NEAR, describes manual review of its Rust contracts and formal specification and verification of its market contract. Certora’s assessment page identifies the proven properties and discovered issues as subjects of the full report.
Other blockchain teams have pursued AI-assisted verification
In May, Aptos Labs described support for AI-assisted formal verification involving dynamic dispatch in Move contracts. Its stated workflow uses AI-generated specifications with Move Prover, according to coverage of the announcement.
Ethereum co-founder Vitalik Buterin has discussed a similar combination of AI tools and machine-checked proofs. As crypto.news previously reported, his May comments identified applications across Ethereum’s roadmap, including consensus, quantum-resistant cryptography and zero-knowledge systems.
Buterin cautioned that proving one layer does not establish that an entire system behaves as intended. His discussion covered verification from high-level specifications through the lower-level implementation.
NEAR’s separate security research has addressed quantum-resistant signing. Near One’s May proposal selected FIPS-204 signature support, using the ML-DSA standard approved by the U.S. National Institute of Standards and Technology.
The May proposal described a testnet target for the end of the second quarter of 2026. Near One said wallet software, hardware wallets, endpoints, and APIs would require coordinated changes to support the new signing schemes.