01
Coverage before commitment
Every model shows readable concepts, method families, provenance, and rating status.
Pre-training dataset diagnostics
Run real, unrated interpretability reads before training.
Proba's technical alpha reads chosen/rejected pairs against version-pinned interpretability primitives, then reports raw direction, coverage, per-family evidence, exclusions, and runtime failures. It does not estimate confidence, corroboration, or training outcomes.
02
Family evidence stays separate
Raw per-family reads remain individually visible; alpha does not collapse them into confidence or corroboration.
03
Reproducible inputs
Reports bind to immutable primitive keys and explicit catalog and policy versions.