Model Provenance Badges

A compact, glanceable badge system that tells an operator at a glance whether an AI output is trained, validated, calibrated, and running on healthy inputs.

Provenance you can read in one glance

The full provenance card answers everything but takes attention to read. During an upset an operator needs the summary instantly, so each AI-derived value also carries a small badge encoding its trust state. The badge is deliberately coarse — it exists to tell the operator whether to lean on this number or scrutinize it, not to replace the detail one click away.

Badge dimensions

A value with all four green is a value to act on. Any amber or red dimension is a cue to open the provenance card and understand the limitation before trusting the number. Out-of-distribution is the most important flag: it means the machine has entered a regime the model was never trained on, which is exactly when an unbadged AI is most dangerous and most confident-looking.

Badges follow the value everywhere

The same badge rides with a quantity across the dashboard, the 3D overlay detail card, the mobile view, and incident replay, so trust state is consistent no matter where an operator encounters the number. In replay, badges are historical — you see the trust state as it actually was at that instant, which is how post-event reviews catch cases where a stale-but-green-looking value misled a decision.

def badge(estimate):
    return {
      'validated':   estimate.model.vv_passed,
      'calibrated':  estimate.calibration_ok,
      'inputs_ok':   not estimate.any_input_imputed_critical,
      'in_dist':     estimate.ood_score < OOD_THRESHOLD,
    }  # 4 booleans -> green/amber/red glyphs, detail one click away

Badges are the machine-readable face of the trust contract; the reasoning behind them lives in confidence and provenance display, and the decision they inform is covered in trust and override UX.