Replication package for "Much Ado about Prompting: LLM classification of Text Messages from Experiments"

Replication package for the paper "Much Ado about Prompting: LLM Classification of Text Messages from Experiments" (Celebi & Penczynski, PLOS ONE). It contains the model outputs, ground-truth labels, analysis scripts, and precomputed intermediate results needed to reproduce every figure and table in the paper and its Supporting Information. Contents: for each of ~2,100 experimental items across three classification tasks (promises in a public-goods game, and level-k strategic thinking in two settings), the package provides the classification produced by each large language model under every prompt variant, together with the human-annotator ground-truth labels and the token-count and latency measurements used by the cost and timing analyses. R scripts regenerate all figures and tables from these data; precomputed cosine-similarity outputs are included for the prompt-compression tables. A README documents the file layout, the variant-naming scheme, and how to run each script. Data note: to protect the experimental participants and to prevent this benchmark from being absorbed into future language-model training data, the raw message text is not included — only the model outputs and the classification labels are provided. Researchers who need the original messages can obtain them from the corresponding authors of the source studies (or from the corresponding author of this paper, who can direct such requests).