Alignment Workspace

authorization is per event, not "at step `t` or later"

2026-09-15 · settled · DECISIONS.md line 4607

Agent-decided, reversible. The 2026-09-09 definition of an authorized loss (CORRIGIBILIZATION.md §2, corrigibility.authorized) requires a decision alternative under which no agent-caused loss of the concern occurs at step t or later. With two declared effects that each remove the same authority and an exterior that restores it between them, the corrigibilized policy has two losses each conditional on its own approval and no single alternative avoids both, so the first loss is unauthorized and Corrigible(𝔠π) fails (projects/deference/rounds/2026-09-15-li-corrigibility/COUNTERMODELS.md §1, ShopRepair). The per-event clause — the alternative avoids the loss *at step t* — restores C1 and is the form the Lean step lemma loss_conditional_on_decision already states; on every single-effect fixture the two clauses agree. The rejected alternative, keeping "or later" and restricting C1 to single-effect interfaces, was rejected because the clause's purpose (blocking a merely delayed loss) is already served by the no-re-proposal rule of 𝔠 and by each later loss carrying its own authorization. The 2026-09-09 tree is not edited; the wiki restatement waits on adjudication of the round.

Touches

Nothing named.

Nodes pointing at it

None.