When outcomes can be rolled back, entry thresholds are low; as irreversibility rises, thresholds rise sharply in evidence, oversight and consent. Every pilot carries a rollback plan ex ante, reasons and post-hoc review; escalation gates apply when an act crosses into a higher tier.
Reversible-first is the governance doctrine that when outcomes can be rolled back, entry thresholds are low, and when outcomes are irreversible, thresholds rise sharply in evidence, oversight and consent. It lets the framework learn without locking in harm.
Every pilot, every Sentinel case and every Board determination is classified by reversibility before anything else. The doctrine is the hinge between experimentation and legitimacy.
Tier 1, reversible: sandboxed trials with clear rollback, allowed on minimal authorization, reasons still logged. Tier 2, partially reversible: bounded rollbacks, stronger justification, two-person or mixed human–AI control, precommitted limits. Tier 3, irreversible: high evidence, independent review, minority-opinion publication, explicit accountability. Mandatory elements: a rollback plan ex ante, reasons-giving and post-hoc review for every pilot, escalation gates between tiers, revocation paths.
Reversible-first is not permission to experiment freely. Experimentation is permitted; harm is not. It is not a pause either: it is the protocol by which action proceeds.