YAI Guard — Prove-It Code Runner

AI coding agents are useful only when their work can be trusted. Today, an agent can claim “I ran it and it passed” without proving what actually executed, where it ran, or whether the evidence was altered. YAI Guard closes that gap. YAI Guard is a production-style web control plane for verifiable agent execution. A user submits a coding task; the app plans it with Vultr Serverless Inference, applies a deterministic policy, and dispatches each command into a fresh Docker/gVisor (runsc) sandbox. The container has a read-only root, a writable workspace, no network, and strict CPU, memory, PID and timeout limits. When the task ends, the sandbox is destroyed and cleanup is confirmed. Every step is recorded in a hash-chained trace: plan, tool call, policy decision, exact argv, exit code, output hashes, file changes, approval, and cleanup. At the end, an anchor seals the event count and final hash, with optional HMAC or Ed25519 signatures. Reviewers can verify the result offline and see a clear level: VERIFIED, CONTENT_VERIFIED, PENDING or INCOMPLETE. YAI Guard never treats “completed” as proof unless cleanup was confirmed and the evidence chain is intact. The app and controller run on separate Vultr Cloud Compute VMs, so the web layer never executes user code directly. Vultr Serverless Inference powers the agent; the controller owns the sandbox runtime. The system includes a durable SQLite task queue, leases, crash recovery, worker fencing and reconciliation, making it a real control layer rather than a demo. YAI Guard is aimed at AI coding platforms, CI/security teams and hackathon judges who need a receipt for every agent run. The public demo shows a real task, its trace, and an offline-verifiable evidence bundle. The code is MIT-licensed and documented for local and Vultr deployment.