Label Quality and Provenance

Labels are versioned artifacts with recorded source and confidence, because a mislabeled disruption or off-normal event teaches a safety model exactly the wrong lesson.

Wrong labels are worse than no labels

Supervised models are only as correct as their labels. A breeder pulse mislabeled as non-disruptive teaches a precursor detector to ignore a real precursor; a burner transient mislabeled as nominal teaches a monitor to stay quiet during an off-normal event. Kronos treats labels as versioned artifacts with explicit provenance and confidence, never as ground truth taken on faith.

Labels come from several sources of differing reliability: automated derivation from physics signals, expert human annotation, and consensus across multiple annotators. Each label records its source, the labeler, a confidence, and any disagreement. Low-confidence and disputed labels are surfaced for review rather than silently used, and their influence in training can be down-weighted.

Label metadata

• Source: automated rule, single expert, or multi-annotator consensus
• Labeler identity and, for humans, expertise weighting
• Confidence and inter-annotator agreement
• Revision history — labels can be corrected, never overwritten silently
• Link to the exact pulse segment and features labeled

label = {
  'pulse':'B-2032-0187','segment':[t0,t1],
  'value':'off-normal','source':'consensus',
  'annotators':['phys-07','phys-12','phys-19'],
  'agreement':0.67,'confidence':0.8,
  'rev':4,'supersedes':'rev:3'    # append-only correction trail
}

Rare, safety-critical classes get the most annotation care: breeder disruption onset times and burner off-normal boundaries are annotated by multiple experts because a one-cycle error in the labeled onset shifts what a precursor detector learns to fire on. Automated rule-derived labels are treated as provisional until an expert confirms or the physics twin corroborates them.

Label provenance is what makes an incident postmortem tractable: when a model misses an event, the first question is whether the training labels were correct. Because labels are versioned, a label correction is itself an event that can trigger retraining of every model that consumed the old label, discoverable through data lineage.