Blast Radius Zero: Sandboxed CSV Fixer

NightMoose built Sandboxed CSV Fixer for data teams handling messy CSV files. Upload a CSV and an optional contract YAML. Vultr Serverless Inference proposes a pandas repair script; a FastAPI control plane on Vultr Cloud Compute runs it in a disposable Docker sandbox. DataContractGate validates the output before a download is offered. Validation failures can feed back into a retry loop, up to three attempts. The reviewed demo shows a real six-row repair and download, followed by a deterministic hostile-script containment drill. The drill is separate from model-generated repair: it deliberately exercises blocked network/environment operations. It is not evidence that the model followed a poisoned prompt or refused one. The sandbox design uses network isolation, resource limits, a non-root execution user, dropped capabilities, and per-attempt removal. The inference key is not passed into the sandbox. This is a containment design, not a formal security guarantee or a claim of gVisor isolation. Docker shares the host kernel; the trusted control plane retains Docker-engine access. The aim is to make generated data-repair work inspectable through attempt history, validation reports, and a containment audit, rather than trusting plausible-looking output. Vultr hosts the web application, orchestration, and sandbox execution, while Vultr Serverless Inference powers repair-script generation.