HUNT — Proof for AI-Generated Code

AI coding agents are becoming remarkably capable at implementing software. But a critical gap remains: completing a task is not the same as proving that the original requirement was actually satisfied. An agent can modify a repository, run its tests, receive green results, and declare success while subtle requirement violations remain undetected. Tests written or selected during implementation may simply fail to challenge the exact behavior that matters. HUNT is an independent verification system for AI-generated software. Instead of trusting an agent’s completion claim, HUNT converts the original engineering requirement into independently evaluated verification obligations. It challenges the resulting implementation, inspects observable system behavior and state, and records evidence showing what was actually demonstrated. Our IBM Bob 2.0 demonstration makes this distinction concrete. Bob receives BUG-147: the account-creation API must be idempotent. Repeating the exact same request must never create a duplicate account. Bob works as the coding agent: it inspects the repository, implements the change, and runs the project’s tests. That is where a normal AI coding workflow might end. HUNT begins where the implementation agent stops. It independently evaluates the requirement, repeats the same account-creation request, and checks the resulting persisted state. The requirement is verified only when the observed evidence demonstrates that the repeated operation produced exactly one account rather than a duplicate. HUNT is designed for developers, engineering teams, security teams, and organizations adopting autonomous coding agents. The developer provides an engineering requirement and allows an agent such as IBM Bob to perform the implementation. HUNT then provides a separate verification layer rather than asking users to trust the implementing agent’s own assessment. What makes HUNT different is the separation between implementation and proof. The coding agent is allowed to build. HUNT independently asks whether the requirement was actually satisfied. The long-term goal is an evidence-driven verification layer for autonomous software engineering: AI can build faster, while humans retain auditable evidence for deciding whether the resulting software should actually be trusted.