Builders Instrument decisions as first-class events: log who decided, where the decision happens, and mark your own staff. Report viewed-denominator accept rates and workload-adjusted utilization. Build the reconciliation layer before quoting outcome claims.

Researchers Publish reconciliation methods, not just frozen test sets. Partner for query-log access under governance — the largest evaluative datasets in medicine are sitting unpublished. Treat deployment assumptions as first-class objects of study (per CMU).

Health systems & buyers Stop asking for benchmark scores; every serious vendor clears them. Ask for the funnel: delivery coverage, viewed-denominator adoption, decision attribution, and what the vendor's own rejection log says. A vendor who can't produce these isn't measuring their product.