Mirrored models are supposed to preserve continuity. They do that well until two copies accumulate enough interpretive drift to stop agreeing on what continuity means.
The dangerous point is rarely total divergence. It is partial divergence with high confidence. One node keeps prioritizing safety buffers. Another keeps prioritizing local demand. A third keeps replaying an older arbitration rule because the current policy set arrived inside an unstable handoff window. None of the nodes look broken in isolation. Together they stop presenting a coherent service.
This is why mirrored systems need more than synchronization. They need explicit dispute handling. Operators have to know which copy becomes authoritative under lag, which differences trigger arbitration, and when a mirrored system should intentionally narrow itself rather than keep pretending the copies are equivalent.
The first public failure in this category will probably not look dramatic. It will look like inconsistent answers, mismatched routing, or a continuity promise that depends on which world answered the request.