Correction rate

What it measures

Characters deleted per 1000 characters of net typed text during a writing session. A writer who deletes 75 characters across a 5000-character draft sits at 15 corrections per kchar — the lower edge of the human-typical band.

Red Stet derives the number from the bundle's event stream: sum the deleted-character counts on backspace and delete events (a single-key press counts 1; a selection-delete counts the size of the selection), divide by net typed characters (keystrokes minus deleted characters), multiply by 1000. Measuring characters rather than key presses means one select-and-delete of a whole sentence registers as the revision it is — and it also means a single large selection-delete can move the rate substantially, which a reviewer should know when reading the number. The signal travels alongside the document; the verifier reads it back at review time.

Why it discriminates

Real composition involves planning, drafting, and revising in real time. The cognitive-process model from Hayes and Flower (1980) treats revision as a fundamental component of writing — it happens at the sentence, clause, and character level during composition, not after. Backspaces are the observable trace of that revision process. A writer types "thier" and corrects to "their" mid-word. Drafts the opening of a sentence, deletes the last six words, restarts. Finishes a clause, scrubs backward to swap an adjective. Each of those interactions leaves a backspace in the event stream.

Model output produced by a system and dictated to type lacks this mechanism. The text is finalized before any keystroke fires. Whether the operator pastes the block in or hand-types the model's output character by character, the revisions don't appear — because the revision happened inside the model, not at the keyboard, and the model doesn't emit backspaces. A 5000-character paste lands with zero corrections. A hand-type of model output lands with whatever incidental typos the operator made while transcribing, typically well under five per kchar.

The mechanical reason is simple. Backspaces require second thoughts at the keyboard. Second thoughts at the keyboard require the act of composing to be happening at the keyboard. When composition happens elsewhere and keyboarding becomes transcription, second thoughts disappear from the keyboard surface and migrate into the source environment — which Red Stet doesn't see.

Two writers with very different prose styles produce broadly comparable correction rates because the signal isn't about the prose. It's about the cognitive shape of producing prose. The deliberate stylist who picks each word and the fast-drafter who barrels through and cleans up later both leave dense backspace traces — different patterns, similar densities.

Research history

The cognitive-process model of writing arrived in 1980 with Hayes and Flower's Identifying the Organization of Writing Processes. They studied composition by recording think-aloud protocols and reconstructing the planning, translating, and reviewing loops writers ran while drafting. Revision was a recurrent sub-process at every level — not a final pass. Bereiter and Scardamalia extended the framework in their 1987 book The Psychology of Written Composition, distinguishing knowledge-telling from knowledge-transforming and showing that mature writers revise during composition rather than after.

The work didn't depend on keystroke data — protocols and direct observation carried it. But the framework laid the groundwork: revision is a property of the act, not an editorial afterthought. Once writers moved to word processors in the 1990s, the question became how to capture revision as it happened.

Keystroke logging filled the gap. The Translog tool emerged from the Copenhagen Business School in the late 1990s; Inputlog from the University of Antwerp followed in the early 2000s. Sullivan and Lindgren's 2006 edited volume Computer Keystroke Logging and Writing established the methodology as the dominant way to study writing process empirically. Leijten and Van Waes's 2013 paper Keystroke Logging in Writing Research: Using Inputlog to Analyze and Visualize Writing Processes in Written Communication codified the approach. Wengelin's 2007 work on pauses and transitions tied keystroke gaps to revision points, showing pauses cluster around the moments writers reorganize a sentence — and backspaces follow.

In parallel, keystroke dynamics as a biometric signal grew out of typing-performance research (Salthouse, 1986) and continuous-authentication work in the 2000s. Modern free-text continuous-authentication systems (Ahmed and Traore, 2014; Vural et al., 2014) include correction behavior — not just timing — as a feature. The field today sits at the intersection of writing-process research and behavioral biometrics: keystroke logs are an established methodology, and correction rate is one of the more discussed features in both literatures.

Key papers

• Hayes, J. R., & Flower, L. S. (1980). Identifying the Organization of Writing Processes. In Gregg & Steinberg (eds.), Cognitive Processes in Writing. Erlbaum.
• Bereiter, C., & Scardamalia, M. (1987). The Psychology of Written Composition. Erlbaum.
• Salthouse, T. A. (1986). Perceptual, cognitive, and motoric aspects of transcription typing. Psychological Bulletin, 99(3), 303–319.
• Sullivan, K. P. H., & Lindgren, E. (eds.) (2006). Computer Keystroke Logging and Writing: Methods and Applications. Elsevier.
• Wengelin, Å. (2007). Pauses, transitions and analytic units in writing. In Torrance, Van Waes, & Galbraith (eds.), Writing and Cognition. Elsevier.
• Leijten, M., & Van Waes, L. (2013). Keystroke Logging in Writing Research: Using Inputlog to Analyze and Visualize Writing Processes. Written Communication, 30(3), 358–392.
• Ahmed, A. A. E., & Traore, I. (2014). Biometric Recognition Based on Free-Text Keystroke Dynamics. IEEE Transactions on Cybernetics, 44(4), 458–472.
• Vural, E., Huang, J., Hou, D., & Schuckers, S. (2014). Shared research dataset to support development of keystroke authentication. IEEE International Joint Conference on Biometrics.
• Conijn, R., Roeser, J., & van Zaanen, M. (2019). Understanding the keystroke log: the effect of writing task on keystroke features. Reading and Writing, 32, 2353–2374.
• MacKay, D. G. (1982). The problems of flexibility, fluency, and speed-accuracy trade-off in skilled behavior. Psychological Review, 89(5), 483–506.

Confidence level

Moderate.

The cognitive-process research underpinning the signal is well-established. Hayes and Flower's framework has 40+ years of replication and extension across languages, genres, and writer expertise levels. Keystroke logging as a methodology for studying writing process is the dominant approach in writing-research labs and has its own decade-plus of refinement through Translog and Inputlog.

What's less mature is the use of correction rate as a discriminator between human-typed and machine-generated text. The keystroke-biometrics literature treats correction patterns as one input feature among many for continuous authentication — useful, not load-bearing. Head-to-head studies comparing correction rates in genuinely human-composed text against operator-transcribed model output are recent and sparse. The signal is meaningful, and the threshold band (15–30 / kchar for human-typical English composition) is grounded in the writing-process literature. But the discriminator framing — as opposed to the original "study the cognitive process" framing — is newer and the calibration evidence is thinner.

Correction rate is also easier to spoof than cadence variance. A motivated adversary can manufacture typos and corrections while transcribing model output; the keystroke trace will look closer to human range. The signal carries weight in the composite, but not the most weight — that goes to cadence variance, which is harder to fake convincingly.

Known limitations

Correction rate has several failure modes a reader needs to hold in mind when interpreting a flagged moment.

• Highly practiced writers produce low correction rates on familiar material. A journalist filing a beat story, a paralegal drafting a routine motion, a teacher writing parent emails — practiced writing in well-trodden formats can sit at 5–10 corrections per kchar without indicating anything other than fluency.
• Anxious writers produce very high correction rates without genuine revision. Some writers backspace nervously — typing a word, deleting it, retyping the same word. The keystroke trace reads as heavy revision; the prose reads as identical to a first-pass draft. High correction rate without revision is real human behavior, not a signal of anything else.
• Pre-planned writing has lower correction rates than exploratory composition. A writer with notes in hand or a rehearsed argument types closer to transcription. Correction rate drops without the writer becoming any less human.
• An adversary who types model output with manufactured typos and fixes can boost correction rate artificially. This is the most direct spoof. Hand-typing model output while inserting typos and backspaces produces a keystroke trace that approximates the human band. Cadence variance and pause distribution are harder to fake convincingly across a long document, but correction rate alone can be manufactured.
• Voice-to-text dictation has no backspace pattern at all. A writer dictating into a microphone produces text without keystrokes, full stop. The signal is absent — and under the current scoring, absence scores 0 on this signal (weight 0.20), which does penalize dictated documents against the hand-typed baseline. The honest reading for a reviewer handling a dictated document is that this signal is not applicable, whatever the number says; a modality-aware fallback (scoring absence as neutral when other dictation markers are present) is planned, not shipped. See Limitations → Assistive input modalities.
• Some writers revise via selection-and-replace rather than backspace. The double-click-to-select-word, then type-to-replace pattern emits a small number of keystrokes for a large revision. Backspace counts under-represent the revision effort for these writers. Red Stet picks up selection-paste-replace events separately, but the correction-rate signal alone reads them as low.
• The 15–30 / kchar band is calibrated on English. Writers in non-Latin scripts, in agglutinative languages, in input-method-edited East Asian scripts — all of these can produce different keystroke-to-character ratios for structural reasons. The thresholds need separate calibration per script and aren't yet validated outside English.
• Short passages produce noisy rates. A 200-character message with zero backspaces sits at 0 / kchar; the same writer over a 5000-character document might sit at 22 / kchar. The rate doesn't stabilize until roughly 1500 characters. Current behavior: the verifier scores this signal at its full 0.20 weight regardless of document length — there is no length gate yet. Readers of short-document scores should treat this signal as advisory themselves; an automatic length-aware downweighting is planned, not shipped.

How Red Stet uses it backspaceScore

The verifier sums deleted characters from backspace and delete events in the bundle, divides by net typed characters (keystrokes minus deleted characters), and multiplies by 1000:

The anchor is 15 / kchar = full human score. Anything at or above the lower edge of the human-typical band scores 100. Zero backspaces across a long document scores 0. The interpolation between is linear.

The score contributes 0.20 to the composite — the second-heaviest weight after keystroke cadence variance (0.22). The weighting reflects two things: correction rate is one of the loudest signals in the literature, and it's also one of the more spoofable. Heavy but not heaviest.

The per-moment surface in the verifier's right-hand panel currently surfaces correction activity in the positive direction only: bursts of three or more consecutive backspaces appear in the Key Human column as evidence of typo-and-fix drafting. There is no flag yet for the inverse case — a sustained stretch with no corrections at all. The session-level score captures that absence (zero corrections over a long document scores 0 on this signal), but it does not appear as an inspectable moment on the timeline. A sustained-absence moment detector is planned, not shipped.

One preprocessing detail: the recorder emits separate backspace and delete event types depending on which key the writer pressed; the analyzer folds both into a single correction count so the rate reflects revision activity rather than which delete key the writer prefers. Selection-deletes carry the size of the deleted range. Word-level deletes (Cmd-Backspace) without a selection currently register as a single deletion — a known undercount for writers who revise word-at-a-time.