Sealed Test Room

THE PROBLEM. Toxicologists hear about a new liver-toxicity model on Monday and want it on last quarter's compounds by Friday. Change control and data-confidentiality rules say "not yet" to anything unreviewed, so the new tool stays outside and the data stays locked. Here the untrusted thing is the tool, and the confidential thing is the data that must go in with it. THE SOLUTION. A person walks through seven screens and decides at each one while an AI agent does the work. She picks the approved tool and the study file (preview, SHA-256, "Synthetic study data" badge). The agent opens a sealed room on a Vultr VX1 host: a Microsandbox microVM with --no-net and one allow rule for the Vultr Serverless Inference endpoint. Before any data enters, it tests its own walls from inside: package index refused, Git host refused, inference endpoint answers. Then it asks her: "Walls tested. Send your file in?" Only after her Yes does the file move. The tool is fetched in a separate data-free room, hashed, installed offline in the sealed one, and run. Results come first, next to two public FDA DILIrank 2.0 labels (the tool's output, not a claim of accuracy). She closes the room herself; both rooms are destroyed and verified absent. THE RECEIPT IS THE PRODUCT. One page IT can file: tool and version, the exact code, SHA-256 of every file in and out, the policy verbatim, the live wall-check rows, every room's start, destroy and verified-absent time, her decisions with times, the calls through the one door. Also JSON, schema-validated. Every failure path ends in rooms proven absent or a loud "not verified". FOR WHOM. Buyer: managers of regulated-data teams (pharma safety, CROs, consultancies). Approver: IT or QA. Pricing hypothesis: per sealed run. Vultr VX1, Microsandbox, Serverless Inference (glm-5.3); 335 tests; MIT. TRY IT. Judge access code: 68neJWDu9CN4RB2M2MkdlRKQ. If sslip.io does not resolve, http://45.63.20.61 serves the same app. Demo data is synthetic.