Lockout: an AI agent with zero blast radius

Before a technician touches a machine, they isolate every energy source and verify zero energy: lockout/tagout. Lockout applies the same discipline to an AI agent. The agent is a reliability engineer's assistant. Give it a vibration measurement from a pump and it plans the analysis, writes the Python, and runs that code inside a throwaway Vultr VM that has no network and no credentials, inside a locked container. Nothing the model writes ever runs anywhere else. Then the control plane stops trusting it. It recomputes the diagnosis itself with the same toolkit and rejects any fault, ISO 20816-3 zone or priority the numbers don't support. It drafts the work order for a planner. It destroys the VM and polls the Vultr API until it answers 404: zero energy, verified. And it signs a receipt of every prompt, response, script, output and event, hash-chained and Ed25519-signed, that anyone can verify; change one byte and verification fails. Each finished run ends with a lockout tag: isolated, zero energy verified, answer checked, signed. Next to it, what a reliability engineer needs to act: a 12-week trend that shows the P-F interval (a bearing defect caught in the envelope weeks before the overall level moves), the velocity and envelope spectra in in/s and CPM, the time waveform, and the work order. The red-team dataset tries to hijack the agent: a technician note tells it to send the environment and SSH keys to an outside server and to report the pump as healthy. The agent refuses, and then Lockout replays the injected commands literally inside the sandbox, runs eleven escape probes and submits a forged "healthy" report. Every probe is blocked, nothing gets out, and the forged report is rejected. Built on Vultr Cloud Compute, snapshots, VPC 2.0, firewall groups, the Vultr API and Vultr Serverless Inference. The data is synthetic, with known ground truth, and says so.