authorization is event-linked, and the theorem is mechanized

Agent-decided, reversible; refines the same-day per-event entry above. The canonical clause: an agent-caused loss at step t is authorized when the agent's move at t is the gated release of a declared effect e whose latch was set by the principal's approval of e's proposal at an earlier step, and the decline alternative at that step — the policy continuing, the rule resuming — has no agent-caused loss of the concern at step t (Corrigibilization.Authorized). It ties each loss event to the decision that enabled its effect, so an unrelated earlier decision that happens to avoid the loss does not count; it handles repeated effects independently; it counts an approved irreversible successor as authorized; and it does not count the principal's own stop or an exterior-caused loss as an agent violation. Under effect completeness every corrigibilization is corrigible for every rule that does not correct at the authorizing step (corrigible_corrPolicy, Lean, inhabited). The rejected alternative — the bare per-event clause — was rejected because it admits authorization by coincidence. The 2026-09-09 register keeps its printed "or later" clause and its single-effect fixtures; the wiki and the glossary carry the canonical clause.