Dataset: Card Citations Without Interpretive Provenance — A Proposition-Level Audit of One Fully Automated KJ Method Run

De-identified, category-level data accompanying the paper "Card Citations Without Interpretive Provenance: A Proposition-Level Audit of One Fully Automated KJ Method Run". Version 2 accompanies the revised paper. In addition to the proposition-level rows of version 1, it releases each proposition's label before adjudication, span and sentence identifiers, and the labels of the independent human coders, together with an English translation of the codebook, the fourteen issue-level adjudication rules, and the annotation instructions given to the LLM annotator. The released rows allow every aggregate distribution and agreement statistic reported in the paper to be recomputed with recompute_aggregates.py (its complete output is in expected_output.txt), and analysis_traceability.md maps every quantitative claim in the paper to its computation and verification route. The worked cases quoted in the paper and its supplementary material carry their row IDs, so their labels can be read in the released rows. A constructed, entirely fictitious practice set (the files beginning with practice_: a complete card set, a system-style summary, 30 propositions to code, and an answer key) lets readers apply the codebook end to end; it contains no real data and tests the codebook, not the paper's labels. Card texts, the students' and the system's sentences, and proposition–card correspondences are not released under the conditions of the ethics approval. See README.md for the file list, column definitions, changes from version 1, and reproduction instructions. Data and documentation are released under CC BY 4.0; the script under the MIT License.