UncertainChart: Data, Specifications, Checker Results, and Repair Outputs for LLM-Generated Uncertainty Visualizations

This archive contains reproducibility materials for the manuscript "UncertainChart: A Semantic Feedback Framework for Repairing Uncertainty Failures in LLM-Generated Data Visualizations." The study evaluates a targeted semantic failure in LLM-generated Vega-Lite charts: when tabular datasets contain explicit uncertainty fields, such as confidence intervals, margins of error, or lower/upper bounds, generated charts may omit, hide, or incorrectly encode this uncertainty. The archive includes processed public-data tables, chart-generation task files, prompts, generated Vega-Lite specifications, deterministic checker outputs, deterministic repair outputs, render-validation summaries, human-audit aggregate files, figure files, validation scripts, tests, and manuscript-support materials. The benchmark covers CDC PLACES confidence-interval data, American Community Survey margin-of-error data normalized to intervals, and Berkeley Earth temporal temperature-uncertainty data. It evaluates five model-access routes and reports checker-level uncertainty communication failures before and after deterministic repair. The archive intentionally excludes local environment files, API keys, cache directories, raw model response text, and raw HTML render directories.