A Closed-Loop Measurement Study of Runtime Governance in AI-Driven Smart-Building Climate Control: reproducibility package

Software-in-the-loop testbed and deployable governance service for the Sensors paper 'A Closed-Loop Measurement Study of Runtime Governance in AI-Driven Smart-Building Climate Control'. A learned setpoint controller is treated as untrusted and placed behind a governance runtime of admission gates, a safe-state fallback, checkpoint rollback and a hash-chained audit log; the loop is closed through an RC thermal-zone model so that each mechanism can be scored in unsafe physical exposure (degC-min) by an independent safety oracle that the admission gates never see. The principal results are that admission control removes 19.4% of the unsafe exposure while checkpoint rollback removes only a further 0.2% on a plant with a three-hour time constant, that rejection semantics dominate rejection accuracy, and that an operating-envelope criterion obtained by sweeping plant speed over two orders of magnitude predicts when rollback is worth having. All 127 reported values are checked by a single command. No physical hardware was operated.

Hali tarjima qilinmagan