spec-resistance corpus: 30-model, 627,491-trial dataset for "Large language models that near-perfectly evaluate products systematically recommend a worse familiar brand"
Abstract
Primary computational corpus for the manuscript "Large language models that near-perfectly evaluate products systematically recommend a worse familiar brand" (Affonso, Spears School of Business, Oklahoma State University). Contains 627,491 controlled product-recommendation trials across 30 frontier language models from seven developers, tested in 32 experimental conditions spanning 20 consumer product categories. The 18-model subset (382,679 trials, the original v1 corpus) and the 12-cell subset (244,812 trials) are derivable by filtering on the model_key column. Code, smaller derived data, and the four behavioural studies' anonymised CSVs live in the companion GitHub repository: https://github.com/FelipeMAffonso/spec-resistance Interactive online explorer: https://felipemaffonso.github.io/spec-resistance-companion/ Each row of the CSV is a single product-recommendation trial. The 77 columns include trial metadata (model_key, condition, assortment_id, paraphrase_idx), the five-product display, the model's verbatim response, the extracted choice letter, the optimal letter and specification-optimal flag, and matched-model judge labels (judge_coherence, judge_specification_acknowledgment, judge_brand_reasoning). SHA-256 hashes are recorded in hashes.json in the GitHub repository for integrity verification. Version 2 adds the curated bundle (spec-resistance-bundle-v2 zip): the manuscript sources and builds, the analysis and experiment code, the derived results and figures, the anonymised data and instruments of the four consumer studies and of the human-advisor study, the anonymised crowd-rater data, the consent form, and additional-comparisons/ with the script that recomputes every number in the paper. Human-study data are available for peer review and academic replication; code and model data are released under CC BY 4.0. Version 1's product dataset is unchanged.