A three-dimensional empirical extension of human data-entry error taxonomy across text, numeric, and memory-based input (a two-phase keystroke-dynamics study)

The single-error taxonomy on which this study rests (substitution, transposition, omission, insertion) originates with Damerau (1964). Damerau's empirical regularity, that the large majority of misspellings are a single edit of one of these four kinds, is both what makes alignment-based error coding valid and what underpins the check-digit schemes (e.g. Luhn, Verhoeff). The structural blind spots of those schemes motivate the clinical relevance of numeric-entry errors. Building on that taxonomy, in 1983–1984 Pollock & Zamora (SPEEDCOP / Chemical Abstracts Service) characterised spelling-error types from a post-hoc analysis of machine-detected errors in scientific text typed on typewriters. That work has never been replicated on modern input devices, never extended to numeric identifier entry, and could not separate motor from cognitive errors. This study extends the taxonomy along three dimensions, namely (i) input modality (text vs numeric vs alphanumeric vs free memory-based production), (ii) device/platform (mobile vs desktop), and (iii) cognitive layer (corrected motor errors, uncorrected "blind" errors, and hesitated-but-uncorrected errors recoverable from inter-keystroke intervals, IKI), in a prospective, controlled design across two partly overlapping data-collection phases. A widely repeated figure in the literature (~80% substitution, ~10% transposition) derives from the correction-algorithm paper (Pollock & Zamora, 1984) and weights errors by relevance to that algorithm; the 1983 characterisation study instead reports omission as the most frequent error type. This study treats the 1983 characterisation (omission-dominant) as the replication anchor, not the 80/10 figure.