Materials for "What a reported LLM-human agreement coefficient does not tell you: a decomposition, five ablations, and four things to report"

Instrument, item set, rater responses, model outputs, and analysis code for the manuscript of the same name, submitted to Behavior Research Methods. Every coefficient, table, and figure in the paper is reproducible from these files. Contents. The 36-item check set with the highlighted motion clause and the pipeline's code for each item; both versions of the browser coding instrument; item-level responses from nine human coders in three arms (v1, v2, and a non-blind pair); outputs from five commercial models across the main arm and four ablation arms (excerpt-only context, three codebook manipulations, the deleted-uncertain-option arm, and repeated runs for self-consistency); and the three Python scripts that compute every reported quantity. Reproducing the paper. python3 scripts/23_p3_tables.py prints every number in the paper's tables; 24_p3_figure.py draws Figure 1; 21_p3_predictions.py evaluates the three registered predictions. Python 3.9+ and, for the figure, matplotlib. No other dependencies. See README.md for the full file map. What is not here. The dream corpus itself is a private personal journal and is not released: doing so would breach third parties' privacy and destroy the contamination-free property the study depends on, namely that the corpus is unpublished and so cannot be in any tested model's training data. The 36 excerpts are the exception, released because the study's claims cannot be checked without them. They appear exactly as the coders and models saw them, carrying the pseudonymisation applied to the corpus at ingest: third parties appear only as bracketed role labels. The mapping from labels to real people is not published anywhere. Coders are identified by letter, as in the paper. Licence. Data and materials CC BY 4.0; code MIT. See LICENSE.txt.