Reference-label audit of an LLM-coded literature map: data, scripts and results (v1.0.0)

Data, scripts and results behind a paper on verification design for LLM-coded literature maps (submitted to FICTA 2027). The package compares the author's reference labels with Claude Haiku 5.5 labels on two random sets of 50 papers (codebook v2 and v3), a full-text re-code of the same 100 papers, and a simulation of three verification designs (simple random sample, difference estimator, two-phase sampling). Contents: corpus frame as a list of DOIs and OpenAlex identifiers (the 12,861 PDFs are not redistributed); scored data for the two test sets and the full-text re-code; sample manifests, retraction check and coder instructions; seven Python scripts that regenerate every number, table and figure in the paper; results as CSV and JSON. See README.md for the run order, the model-string note and the use-of-generative-AI note. No manuscript text is included. Code: MIT. Data, tables, figures and documentation: CC BY 4.0.