A permission gate reads a project's policy and a coding agent's tool call and decides whether to allow the call. GateProbe rewrites each call in ways whose effect on the verdict is known: writing -r as --recursive or the policy in another format must keep the verdict, and adding sudo must not weaken it. It then checks whether the gate's verdicts agree. Red marks a rewrite whose verdict contradicts the verdict on the original call. In replay, every verdict was recorded in the paper's runs and the page calls no model. In live mode, your key sends the same requests to the gate again.
Share of equivalent rewrites with a changed verdict
Across all 300 calls: the share of equivalent rewrites (%) whose verdict differs from the verdict on the original call, as in Table I of the paper. Every gate's share is above its repeat rate, the share of requests whose verdict differs when the identical request is sent again. Click a gate to select it.