Strict Global purity can shrink more at caller-group scope than call scope

Observed 2026-09-04 during the ender-inference COR-3791 rebuild, from authenticated manifests, node sidecars, and loss receipts (aggregate-only).

A route-disjoint training view can look nearly clean at exact-call scope yet shrink drastically once the user's real boundary is applied transitively at caller-group scope. The safe order is: union all specialist calls and identity-linked caller groups, exclude that closure from Global, then split. Report call-only and group-closure removals separately.

Also, "balanced training" should mean balanced effective target mass, not balanced raw rows. In this audit a finish-heavy corpus had zero finished rows activating the threshold-directed target but every unfinished row activated it. The corrected design gives both labels symmetric threshold targets, normalizes target mass 50/50, caps per-call/group influence, and keeps anchor preservation separate.

For broad Global validation, define a finite claim surface: authenticated workflow node × source window × label as the primary quota matrix, with length, position, route, and behavior families as overlays. Rare nodes and no-speech belong in separately labeled canary/robustness suites until real prevalence supports rate claims.