KeepCount: A Stage-by-Stage Audit of What Images of Behavioral Data Keep, with Keystroke Tiles as a Worked Example

Works that turn behavioral data into images summarize it in stages, yet what an image keeps is rarely counted. KeepCount is an audit procedure: cut the transform into stages and, at each stage and against comparisons fixed before the count, count how well it tells one person from another and whether an operation on the data shows, with positive controls and, for a yardstick narrower than the person's own variation, a null of exchangeable gaps. It is applied to a transform that turns keystroke times into shaded tiles, one per sentence, through speed and rest, with three public keystroke datasets. This is the authors' original manuscript, submitted to Information Visualization (Sage) on 2026-11-02. It has not been peer reviewed. Supplements A-E and supplement-cells.csv are the versions submitted with it. The conditions, code, records and git history are archived at https://doi.org/10.5281/zenodo.23858736. The use of AI tools is stated in the manuscript. The keystroke data are not redistributed. KUPA-KEYS (Velentzas et al., LREC-COLING 2024) is released under CC BY-NC-SA 4.0, 136M Keystrokes (Dhakal et al., CHI 2018) for non-commercial use with attribution to its authors, and the essays of the Kaggle competition Linking Writing Processes to Writing Quality under the competition's rules; the article reports only counts and summaries over persons computed from them, not their contents (the minimum and maximum of a summary are single persons' values).