NEAR AI Sets Out Verification Goal; Math Benchmark Is Not a Network Security Audit

NEAR’s 16 October explanation connects AI-assisted proof research to a protocol-verification goal. Its reported mathematics benchmark is evidence about proof generation, not certification of deployed blockchain software.

Published 19 October 2026 · Source explanation: 16 October 2026 · NEAR / AI research

A research goal, not a completed protocol audit

NEAR AI wants to make formal verification more practical and ultimately apply it to the NEAR protocol and core contracts, according to NEAR’s 16 October explanation. The announcement describes a direction of research; it does not establish that the live network has now been comprehensively verified.

Formal verification checks precise requirements within a mathematical model and stated assumptions. AI can search for candidate proofs while a proof assistant checks the result. The guarantee remains tied to the requirement actually expressed and to how faithfully the model represents deployed code. Testing and code review still have a role.

What the benchmark measures

The NEAR AI research repository reports 672 machine-validated Lean 4 solutions: 670 solvable upstream PutnamBench problems and corrected versions of two defective formalizations. Those qualifications matter when interpreting a headline total. These are mathematics problems, not a demonstrated verification of NEAR’s production contracts.

The repository describes checks against incomplete proofs and altered theorem statements, plus additional semantic validation. NetNapz reviewed that description but did not rerun the experiment. Its reported $111.85 cost uses prices applied to the submission, not a fresh October quotation or a complete software-verification budget. The longest problem took more than three days. Cheap inference therefore does not necessarily mean quick completion.

NetNapz assessment: follow the production evidence

Assessment: the investment-relevant question is whether research becomes a maintained engineering practice. A useful next disclosure would identify the production-code version examined, properties proved, assumptions, uncovered components and reproducible evidence. Without that bridge, a mathematics benchmark should not be marketed as a network-wide security assurance.

For example, proving an accounting rule does not answer every access-control or operational question. Readers should ask which failure modes a particular proof addresses, rather than treating verification as a binary badge. Nor do these research results alone establish increased token demand or a change in NEAR’s market value.

Sources and verification limits

Checked 19 October 2026: NEAR’s dated explanation, its linked research repository and the benchmark maintainers’ leaderboard notes. Benchmark figures are attributed to the researchers, not independently reproduced here. No current ranking, live price or completed production audit is asserted. The existing featured attachment and its attribution were checked; the logo was visually inspected. This report is not investment advice.