TypeAlias.ImportMappingConfidence

type ImportMappingConfidence = ObjectEnum<typeof ImportMappingConfidence>;

What a client should do with a suggested correspondence — the policy applied to its evidence, decided server-side.

Derived from ImportMappingEvidence here rather than left to each caller, so the cut-off has one home: it is calibrated against the evaluation corpus, which lives with the backend, and retuning it must not require rebuilding a client. Clients read this field; they should not re-derive it from evidence or origin.

The three levels are deliberately coarse. With one false mapping across nine samples, the corpus can order the evidence classes and say where to cut; it cannot support finer gradations, and pretending otherwise would dress over-fitting up as calibration.

• high: apply it. The reviewer's attention is better spent elsewhere.
• review: apply it, but say so. Plausible and unconfirmed — the correspondence a reader should look at first.
• low: do not apply automatically. Offer it as a suggestion instead, so the destination stays empty until someone accepts it. Either the values contradict the correspondence, or a required destination is at stake and a silent mistake there fails the import on every row.