Skip to main content

Module Reference

Evidence checklist​

  • decision contract and non-AI baseline
  • provenance-aware dataset card
  • leakage-resistant split with tests
  • reproducible run metadata
  • metrics tied to costs and capacity
  • calibration, uncertainty, slices, and error taxonomy
  • rejected alternatives and stop condition
  • peer reproduction and AI-use declaration

Terms​

Baseline: simplest credible comparator for the same decision.
Leakage: information unavailable at the real decision point entering development or evaluation.
Calibration: agreement between predicted probabilities and observed frequencies.
Slice: defined subset used to inspect differential behavior.
Evaluation contract: frozen population, split, metrics, exclusions, and evidence rules.