The Last Reversible Moment: A Constitution for Consequential AI

This public declaration presents a constitutional and architectural framework for consequential AI centered on the last reversible moment before a model or agent proposal becomes a protected effect. It argues that models may propose, agents may prepare, humans may delegate, and policies may evaluate, but the system or institution accountable for committing the effect must retain a current, receiver-owned acceptance decision before money moves, data is released, authoritative records change, infrastructure is modified, high-stakes recommendations become interventions, or synthetic media is released into contexts where people or systems may rely on it as real.

The declaration defines Action Acceptance as the resolution of one complete exact action against current authority, policy, evidence, declared effect count, remaining authorization capacity, and replay state. It defines the Acceptance Plane as the architectural control function between proposal and protected effect. It establishes three minimum outcomes—ACCEPT, HOLD, and REFUSE—and calls for equivalent, non-bypassable enforcement across every effect-capable route, canonical exact-action representations, effect-owner-defined material dependencies, meaningful recourse, and reviewable decision receipts for both effect and no-effect outcomes.

Its central thesis is that plural intelligence should not require effect blindness. Access is not acceptance. Permission is not proof. Capability to act is not authority to act. Authority may travel, but acceptance is not inherited. As intelligence becomes more plural, the requirement for current, receiver-owned acceptance must remain invariant.

This record is the canonical public declaration, version 1.0.0. It is not a technical standard, certification, legal opinion, implementation authorization, third-party endorsement, or patent license.