A robot was one step from running a synthesis that chemistry says can never work. Deterministic code caught it and flashed red before a single reagent moved. That red card is what we built for. In this self-driving-lab demo, a Gaussian-process optimizer proposes the next lead-free perovskite to try, and every candidate first passes through a digital-twin gate that lives outside the AI. The gate runs five real, published chemistry checks: Goldschmidt tolerance, octahedral factor, charge neutrality, Sn(II) oxidation in humid air, and the organic-precursor decomposition ceiling. Fail one, and the candidate is BLOCKED BY DIGITAL TWIN before any reagent or robot-hour is spent, with the exact rule it broke named on the card. Across one recorded campaign, the gate stopped 38 doomed syntheses before they ran. That prevented $4,419 of reagent waste and 90.1 robot-hours, with zero infeasible experiments executed. Then the scoreboard: the loop met a hard spec in 75 experiments, where random screening on the identical hidden objective would expect around 1,491. About 20 times fewer. One honest caveat. That objective is a physically-motivated synthetic benchmark, not real lab or DFT data, and no real material was synthesized. The value that holds at any model quality is what we lead with: doomed work blocked before it burns money, and a tamper-evident record of every decision the loop made. Agents advise, code decides. In a lab that runs itself, that is how autonomy earns trust. We build the brain, the safety gate, and the provenance layer on your existing hardware. If your team is working out how to make lab automation fast and auditable, we would genuinely like to hear how you are approaching it. #SelfDrivingLabs #LabAutomation #MaterialsInformatics #AIGovernance #PerovskiteSolar