Many systems fail not because execution stops working, but because continuity between operational phases is never globally validated. Each phase — execution, recovery, delegation, detection — can remain locally valid while the system progressively loses authoritative continuity across phase transitions. The rupture does not appear until a boundary event forces global reconciliation. We call this class of failures phase-boundary failures. They are not state loss, not monitoring gaps, and not ordinary distributed consistency problems. They represent a structural condition in which no architectural layer is responsible for validating that local phase validity adds up to global state coherence. The defining property: the system cannot detect its own incoherence because detection operates within phases, not across them. This note defines the failure class, characterises its shared geometry across four observed cases, and distinguishes it from adjacent failure types.
Roman Kir (Sun,) studied this question.