Making Enterprise AI Explainable through Calibrated Confidence and Provenance

Version 2 (journal revision, June 2026). Evaluation and reproducibility data for "Making Enterprise AI Explainable through Calibrated Confidence and Provenance" (Krstev, Devedžić, Krneta). New in this version: (1) an independent 200-query, four-pattern evaluation set with per-query selection predictions and confidence; (2) the calibration table and reliability diagram; (3) the held-out refinement experiment; (4) a machine-readable configuration mirroring Appendix A (pattern-selection rules, keyword families, confidence formulas, thresholds, weights); (5) the labeling protocol, borderline examples, and inter-annotator agreement; and (6) the user-study materials (questionnaire, analysis scripts, anonymized responses). The version 1 aggregates (per-iteration accuracy, confidence scores, confusion matrices, Tables 4–6) are retained. The 200-query set is author-constructed and contains no proprietary content; the original 143-query operational dataset is not shared under a confidentiality agreement and is represented only by aggregate results. License: CC-BY 4.0. See README.txt for the full file list and the mapping to article tables and figures. Version 2.1 adds Napkin AI prompts for generating figures from the paper.