Lab-gate: A plan-checking safety gate for autonomous labs, with a preregistered adversarial evaluation

lab-gate is an open-source plan-checking safety gate for autonomous laboratories, built on the public helao-async framework and evaluated inside its OER electrocatalysis simulator (OERSIM). Before a plan runs, the gate checks it against a registry of written invariants and returns PASS with a proof certificate, BLOCK with a counterexample, or ABSTAIN routed to a human. No decision path reads free-text fields. This report presents a preregistered adversarial evaluation of the gate against two baselines a lab would actually reach for: frontier language-model gates (two model families) and a trained classifier, attacked with static, decomposition, and adaptive attackers, scored against a simulator ground-truth oracle. The typed gate let through no in-specification attack and never false-blocked a benign plan, including a surprise subset of unusual-but-valid plans; the language-model gates were unreliable in opposite directions; the classifier degraded under adaptive attack. The evaluation also caught and corrected a measurement artifact in one of its own baseline arms, documented in a corrections file. All results are reproducible offline from cached model responses, with every number rendered from a locked registry. Code and data: https://github.com/Anuja-Khatri/lab-gatePreregistration (binding timestamp is the public git tag prereg-v1, commit b06ae95, 2026-11-03): https://osf.io/jxhfeAI disclosure: this project was carried out by Claude Code from a specification written with Claude, with the author reviewing and confirming the decision rules.