Skip to content

Advanced Topics

This section provides deep dives into Stickler's internal algorithms and advanced features. These pages assume familiarity with the basics of defining a StructuredModel, running comparisons, and reading evaluation results.

Contents

  • Classification Logic -- How Stickler categorizes comparison results into True Positives, False Alarms, False Negatives, False Discoveries, and True Negatives, and how derived metrics (precision, recall, F1) are calculated.

  • Hungarian Matching -- The optimal bipartite matching algorithm used to pair list elements before classification, including worked examples with List[StructuredModel].

  • Threshold-Gated Evaluation -- How recursive field-level evaluation is gated by a similarity threshold so that only well-matched object pairs receive detailed analysis.

  • Dynamic Model Creation -- Creating StructuredModel classes at runtime from JSON Schema or custom JSON configuration, enabling configuration-driven evaluation without writing Python model code.

  • Confidence Metrics -- AUROC-based confidence calibration: attaching confidence scores to predictions, measuring how well confidence correlates with accuracy, and interpreting the results.

  • Aggregate Metrics -- Automatic hierarchical confusion-matrix aggregation at every node in the comparison result tree.

  • Model Export -- Exporting and importing model schemas in JSON Schema and Stickler-config formats for round-trip serialization, version control, and cross-system interoperability.