The floor test (MF-2026-001): published data, codebook and archive manifest

Data released with "The floor test", a study by Denes Csaszar (Makefield) of how much AI answer engines change their answers to the same questions over time, measured against each question-and-engine cell's own noise floor: how much two answers differ when the same question is asked twice in the same sitting. Up to eighteen questions about one little-documented company were asked of up to eight answer engines in seven waves between 4 July and 27 August 2026. Of the 103 cells with a floor, 85 changed beyond it; 40.9 % would be expected by chance. The study is exploratory and compares wording, not facts. This upload contains the two published CSV files (cells, gap pairs) and the summary JSON exactly as served on denescsaszar.com, with their README and checksums; a machine-readable codebook (Frictionless Data Package) and SPSS (.sav) versions generated from the CSV files and tested to match them; a manifest of the unpublished files by SHA-256; and a plain statement of what is lost or was never recorded (WHAT-IS-LOST.md). Record page, with every number linked to its source: https://denescsaszar.com/ . Which numbers recompute from these files: https://denescsaszar.com/numbers . Version 1; corrections, if any, are published as new versions with a dated note.