A Taxonomy of Hallucination in LLM-Based Form Filling: An Empirical Study Across Models and Domains : Replication Package

Data, derived metrics, and figures supporting Paper 2 of the EchoFill PhD research arc (LLM hallucination taxonomy in structured form-filling). 43,640 field-level predictions across 4 models (Claude Sonnet 4.6, GPT-5.4-nano, Gemini 2.5 Flash Lite, GPT-3.5-turbo), provided to enable independent verification of all quantitative claims in the manuscript. Headline finding: an initial "Claude highest hallucination rate" reading was traced to an evaluation-pipeline artifact. After correction, GPT-5.4-nano and Claude Sonnet 4.6 are statistically indistinguishable on field accuracy (McNemar p = 0.106). Fabrication is rare (177/43,640, 0.41%). See README.md for the full reclassification audit trail and column reference.