← Back to redstet.com
Methodology · Red Stet documentation

How Red Stet measures authorship process

Red Stet's verifier doesn't run a black-box detector. It surfaces specific patterns from a writing session that 30+ years of behavioral-biometrics and writing-process research has characterized in hand-typed composition — and that Red Stet's own calibration work extends toward separating hand-typed sessions from mechanical insertion. The literature provides the within-human distributions; the human-versus-machine thresholds are ours, calibrated on a small corpus and published openly. This is where we show our work — the third-party studies, the confidence levels, and the limitations.

What this section is. One page per signal Red Stet uses, plus five methodology pages that explain how we set thresholds, what the composite computes, what the integrity checks prove, what the system cannot do, and how this approach differs from AI-output detectors. Citations are real. Confidence levels are honest. Limitations are explicit.

Who it's for. Academic-integrity board members reviewing whether Red Stet's evidence is admissible in their process. IT directors evaluating the tool against vendor claims. Journalists writing about the AI-authorship landscape. Researchers who want to verify our reading of the field. Writers curious about what the recording layer actually captures.

What we claim. The composition fingerprint is evidence FOR human authorship when its signals are present. It is not a verdict, not court-admissible biometric proof, and not an AI detector. We are honest about this on every page.

Per-signal methodology

Each of the seven signals Red Stet's composition fingerprint measures, with the research history, mechanical reason, key papers, confidence level, limitations, and how Red Stet weights it.

Keystroke cadence variance
The time between consecutive keystrokes, and how much it varies. The best-validated behavioral biometric — 45+ years of research from Gaines 1980 through modern continuous authentication.
Well-established
Correction rate
Backspaces per 1,000 typed characters. Cognitive process model of writing (Hayes & Flower 1980): revision is intrinsic to human composition, observable as keystroke-level edits.
Moderate
Paste patterns
Paste size, frequency, and source (internal cut-and-rearrange vs external clipboard). Working-memory chunking (Miller 1956) makes paragraph-size external pastes mechanically incompatible with in-flow composition.
Mixed confidence
Mouse path geometry
Curve complexity of cursor paths between rest points. Fitts's law (1954) + Pusara & Brodley 2004 onward — human motor control produces sub-movement curvature; automation produces straight lines.
Emerging
Click patterns
Whether clicks fall on grid-aligned pixels or with sub-pixel motor noise. Modest individually; useful in combination with other signals. Bot-detection literature, less academic.
Experimental
Thinking pauses
Long inter-event gaps. Schilperoord 1996 + 40+ years of cognitive psychology of writing — macro-pauses are the observable traces of cognitive composition work.
Well-established

Methodology pages

How we set investigation thresholds, what the composite score actually computes, what the integrity checks prove, what the system cannot do, and how this approach differs from output-text AI detectors.

Investigation thresholds & baselines
What "flagged" means numerically, how we set the <1% / 1–2% / 2–5% / >5% ladder, and the small calibration corpus our thresholds rest on. Honest about scope.
Combined fingerprint & composite scoring
How the seven signals combine into one number. Weights, formula, verdict tiers, and why no individual signal is enough on its own.
Integrity checks
What the four checks behind "Verified" prove (internal consistency), what they can't (authenticity without the signature layer), and what the recorder captures beyond the scored signals.
What this system cannot do
The typed-out-model-output adversary, short documents, non-Latin scripts, accessibility input, measurement-environment confounds, the aware adversary. Eleven honest gaps named explicitly.
Red Stet vs AI-content detectors
We measure process; they measure output. Different signal, different failure modes, complementary when stacked. Liang 2023, Sadasivan 2023, Weber-Wulff 2023.
Complete bibliography
All cited research, organized by topic. Open-access links where stable URLs exist. The full reading list.

One framing note before you read. Red Stet's product is "is this human-authored?", not "is this AI?" Every signal on every page is described as evidence FOR human authorship — when it's present, the document is consistent with hand-typed composition. When it's absent, the patterns are uncommon in hand-typed documents — never AI-detected.

That framing is locked at the product level (see the 2026-06-08 decision drops). It's not marketing. The signals genuinely don't tell us whether a model wrote the text; they tell us whether the SHAPE of the writing process is consistent with hand-typed composition. A skilled adversary who types model output character-by-character will produce a clean composition fingerprint. We name that limitation explicitly on the Limitations page.

The integrity board, the editor, the reviewer reading a Red Stet verification gets evidence to interpret — not a verdict to apply. That distinction is the whole point of this section.