Pre‑execution oversight is the cornerstone of trusted AI control: a fallible LLM monitor reviews planned actions before they become irreversible. Over‑blocking reduces usefulness and forces deployers to disable the monitor. Every protocol must fix a unit of verification—the number of actions examined per call. Existing designs take this unit for granted and never measure its impact on error‑prone monitors. Natural traces cannot isolate the effect because review length co‑varies with error type and position. Relying on catch alone is misleading: rejecting everything appears to catch all errors. To measure the unit alone we need boundary variation with a matched clean control. We introduce the twin‑prefix framework, which supplies both. Each gold plan yields a prefix with one injected, environment‑accepted error and a clean twin that differs in a single write. Judging each pair at five nested lengths ties verdict changes solely to the unit. Discrimination is scored by pre‑registered informedness, defined as catch minus false rejection. Longer reviews raise catch, but false rejection climbs in lockstep. Informedness peaks at reviewing one or two actions for all six judges across two domains: longer windows make zero‑shot monitors more rejective, not more discriminative. Replaying withheld observations shows the failure is largely due to observation deprivation. Safety cases should state the chosen unit and co‑report the clean series. Our framework is the first controlled, pre‑registered instrument for this choice and never reads catch alone. A calibrated short unit recovers up to 0.95 informedness over an eight‑action review, and no tested label‑blind policy consistently beats it.
Blogger's Review: This work highlights how the length of the supervision window critically shapes LLM monitor behavior and offers a reproducible experimental setup, providing valuable guidance for building more reliable AI safety mechanisms.