Authorized Is Not Caused: Overattribution in the Emerging Agent-Authorization Stack

When an autonomous system acts under valid authority but the world—not the system—causes the outcome, who is accountable? The 2026 wave of agent-governance work answers before the fact, by authorization: an agent must be entitled to act, and that entitlement record becomes the substrate of accountability. This is a genuine advance, but it makes one record carry two independent questions—was the system permitted to act? and did the system actually cause the effect?—and answers only the first. This paper identifies the resulting failure mode: structural overattribution. Wherever an effect is authorized yet externally caused, the record still ascribes it to the system. We show that the downstream literature on responsibility gaps presupposes a causal substrate that authorization-only attribution has already removed, so its gap-closing machinery runs vacuously—blame settles on a non-cause while the true external cause escapes the record. We propose a corrective: a second, ordered gate—self-causation—applied after establishment. We state the two-gate structure precisely and machine-check it in TLA+/TLAPS, showing the overattribution result (Correct ⊊ AOnly) is robust to dropping the model's one contestable assumption. We ground the diagnosis in a documented fatality (the 2018 Tempe autonomous-vehicle case) and in a deployed governance layer that renders a self-causation signal yet never gates on it.