A mandatory middleware layer for high-stakes agent pipelines that requires at least one external reference point before an output can be marked verified. Self-correction loops are explicitly rejected as verification — the system requires a tool call, API response, human confirmation, or database lookup. Outputs that cannot be anchored to an external reference are flagged as "unverified confident" rather than "verified."