Confidence and Provenance Display

For any estimate an operator can see how sure the model is, which model produced it, on what data, and which live diagnostics currently constrain it.

Two questions behind every number

Trust in an AI estimate needs two answers: how sure (confidence) and why should I believe this instance (provenance). L6 answers both on demand for every value on screen. Confidence comes from the twin's uncertainty machinery; provenance is a compact record of the model version, its training/validation lineage, and the specific live diagnostics feeding this estimate right now.

The provenance card

• Model identity and version (e.g. equilibrium PINN, release hash)
• Training-data lineage and validation status (V&V pass/fail, when)
• Live inputs: which diagnostics are contributing and their health
• Fallbacks active: is any input GNN-imputed rather than measured?
• Calibration status: is this model's confidence currently trustworthy?

This makes the difference between a fully-measured estimate and one propped up by imputation immediately visible. If a breeder equilibrium reconstruction is running with three Mirnov coils dropped and their signals imputed by the GNN, the provenance card says so and the confidence reflects it — the operator is never surprised by a degraded input after the fact.

Provenance under diagnostic loss

Fusion diagnostics degrade under neutron and thermal load, so provenance is dynamic. As live inputs drop, the card updates in real time and the corresponding overlay regions shift to hatched/low-confidence. For the burner, if plug-density diagnostics thin out, the ambipolar-potential estimate's provenance shows the reduced constraint set and its confidence widens accordingly.

{
  "estimate": "psi_boundary",
  "model": {"name":"equilibrium-pinn","version":"1.4.2","vv":"passed 2027-Qx"},
  "confidence": 0.71,
  "calibrated": true,
  "inputs": [{"tag":"mirnov_04","status":"ok"},
             {"tag":"mirnov_11","status":"imputed_gnn"},
             {"tag":"flux_loop_02","status":"ok"}],
  "fallback_active": true
}

Provenance is also what makes an override defensible: an operator who overrides the AI can see and record exactly what the AI was working from — see trust and override UX and the machine-readable provenance badges that summarize this card at a glance.