Copilot: three scenarios, one verified base.
Two analysts forecast the same question from different evidence, and a third reconciles them. Before any of them runs, the SAGE-Commons engine grounds the historical series and states what that history supports.
Read the two vocabularies separately. The engine's verdict covers the
history only, and is produced with no language model involved. Everything the
three scenario agents write is model output, carries its own labels
(MODELLED, ASSUMPTION_DRIVEN, DIVERGENT, UNVERIFIED_RECALL, UNSUPPORTED),
and is not engine-verified. A projection is not evidence.