A.19.UINDM:5.3 - Show (U.Episteme) — robust evaluation across environments

A research lead wants indicators for model robustness under distribution shift (different hospitals, sensors, geographies).

• The declared characteristic-space basis includes many candidate metrics (accuracy slices, calibration, subgroup error, OOD detection quality).
• The indicator choice policy is “invariance‑driven”: prefer indicators whose semantics remain stable under environment changes; deprioritize proxy metrics known to be environment‑sensitive.
• UINDM returns an indicator set used by the scoring and comparison stages; uncertain indicators are handled via tri‑state guarding rather than coerced to zero or silently dropped.