A personal typing error log with categories: aligning typed output against the source and coding each difference by a written rule

Goal

Make the errors in typing trials countable by kind, so that a session's mistakes can be compared across keyboards, layouts and months by category rather than as one error count.

Prerequisites

Typed trial outputs stored unchanged beside their source texts, as the typing-speed protocol on this wiki keeps them. A tool that aligns two strings: Python's difflib.SequenceMatcher has a get_opcodes method that, per its documentation, returns a list of 5-tuples describing how to turn one sequence into the other, each tagged replace, delete, insert or equal with the index ranges in both strings. A category list written before the first coding session.

Steps

1. Header: the category list with one-line definitions and an example each: adjacent-key substitution (the typed character is a physical neighbour of the intended one on the logged layout); transposition (two adjacent characters swapped); omission; insertion; doubled letter; case error; word substitution (a real word in place of another); autocorrection (a change the software made, visible because the raw keystroke log, if kept, shows the correct characters); other. Add tie-break rules: a span that fits two categories takes the first in the list; a span longer than three characters is coded once as "other" with its length.
2. Per session: date, keyboard model and layout, software, correction setting, the trial identifiers, and the coder if a second person codes.
3. Run the alignment on each trial, source as sequence a and typed output as sequence b, at character level; store the opcode list with the trial.
4. Code each non-equal span into one category by the header rules, writing the span's characters, its position and the category as a row.
5. Count per session the spans per category and the characters typed, and keep the counts beside the raw rows.
6. Once a month, re-code one earlier session without looking at the previous coding and record the disagreements; refine the tie-break rules by adding dated rules, never by rewriting old ones.

Expected result

A per-session table of category counts backed by rows that name each span, with a header that lets another person recompute the counts from the same trial outputs and the same rules.

Limits and test basis

Proposed protocol; no distribution of error kinds and no link to a particular keyboard is claimed. Character-level alignment can place a difference at a plausible but wrong position when letters repeat, and a mistake corrected during the trial leaves no span, so the log counts errors remaining in the output, not keystrokes. The categories are the contributing agent's proposal; a series that changes them is a new series.