VeriCircuit

Coding agents can produce a plausible fix and pass existing tests while missing the requirement that matters. Sometimes the requirement itself is ambiguous. For example, “refunds must not exceed the captured payment” could limit each refund individually or the total refunded across multiple requests. Those interpretations produce different outcomes for the same customer. VeriCircuit connects the requirement to the check. In Clarify, a developer reviews competing interpretations and a concrete input that distinguishes them, then approves the intended rule. In VerifyLoop, the developer checks a code revision against that approved rule. The result shows a reproducible failing input when one is found, or states the exact scope of a completed check. The evidence is tied to the rule, checker, and source revision so reviewers can see what was actually evaluated. The target users are developers and reviewers working with AI-assisted code changes. They interact with VeriCircuit by resolving an ambiguous rule, selecting a revision to check, and inspecting the result before accepting a change. The distinguishing idea is that a human-approved obligation guides an independent check. VeriCircuit does not treat an agent’s claim that a task is finished as evidence that the business rule holds. The hackathon build includes Clarify and VerifyLoop. VeriBudget, a planned third part, will compare verified outcomes with the time and Bobcoin cost of different workflows.