On November 11, 2026, at 12:36 (UTC+8), an article described AI formal verification aimed at smart contracts. The workflow below is that account, restated as a task you can follow. It is a conceptual description of the reported method, not a menu path from a product that was run for this page.
Task
Infer properties from a contract, turn them into formal rules, and have a prover check those rules. Treat a finished run as one of three results: the stated rules are proved, the check fails with a counterexample, or the result is uncertain. A proof covers only the rules that were actually checked. It does not show that every requirement was written down.
Prerequisites
The description assumes:
- Contract code and documentation the agent can read
- A prover that receives the generated specification and verification rules
- Expert review for complex properties, and for whether the specification left a real requirement out
Steps the article describes
- An AI agent infers properties and invariants from the code and the documentation.
- It generates a formal specification and verification rules.
- Those rules are handed to the prover.
- Read the prover result before treating the contract as checked.
How to recognize the outcome
- Failure: the check does not hold, and a counterexample is given.
- Timeout or insufficient resources: the result is uncertain. That is not a successful proof, and it is not, by itself, a confirmed defect.
- Partial success: only the properties that were proved are covered.
AI code review and AI formal verification
The article contrasts the evidence from AI code review with the evidence from AI formal verification. It says AI can lower the cost of writing specifications and of proof engineering. It also says a generated specification may omit requirements, and that complex properties still need an expert.
What was reported for Certora AutoProver
In an experiment on Aave V4 Hub, Certora AutoProver generated 104 rules and proved some properties. It did not prove share-value monotonicity. The source does not treat that run as proof that every relevant property held.
Use cases, cost, and pricing
The same article lists use cases, cost, and pricing. This rewrite does not add prices or scenario lists, because those details are not stated beyond the fact that the original lists them.
Original article: 智能合约的 AI 形式化验证：成本与使用场景, published November 11, 2026, 12:36 (UTC+8).