Paste patterns

What it measures

Two numbers. The paste ratio is pasted characters divided by typed characters across the session — how much of the final document arrived through the clipboard rather than the keyboard. The paste source split is the share of paste events that came from inside the editor (cut a sentence here, paste it three paragraphs down) versus the share that came from the system clipboard (something the writer copied from a different window, app, or document).

Each paste event in the bundle carries a len field (how many characters arrived), a content field (the literal pasted text, recorded for replay fidelity, with credit-card / API-key / SSN-class material redacted via the sensitive-content gate the writer interacted with at paste time), and a pos field (where in the document the paste landed). The source category — internal cut-and-rearrange vs external clipboard — is derived at analysis time from the temporal and length proximity of a recent cut or copy event: a paste within 5 seconds of either, with a length match within ±20%, is classified as internal; everything else is clipboard. A matched cut is consumed by its paste; a matched copy is not, so pasting the same copied passage several times reads as repeated internal duplication. The verifier walks the event stream, sums those fields, and scores the two ratios against thresholds calibrated on Red Stet's sample corpus.

Why it discriminates

Human writing in flow produces text in chunks the size of working memory. The classical estimate from George Miller (1956) is seven items, plus or minus two — refined downward by later work toward three to five "chunks" for unrehearsed material. In sentence terms that's roughly three to seven sentences before the writer hits a natural break and pauses. A four-hundred-character block of polished prose does not arrive whole. It accretes, with revision, over minutes.

A paste of four hundred characters arrives in a single keystroke. The two events are mechanically incompatible. Either the writer composed those four hundred characters somewhere else and then pasted them in (the editor was not the place of composition), or the writer pulled them from a non-compositional source — a quoted passage, a code excerpt, a previous draft, model output. Each of those reads differently, and the source field is what disambiguates.

Internal pastes — cutting a sentence and moving it to a different paragraph — are intrinsic to human composition. Hayes and Flower's 1980 process model of writing names this explicitly: writers translate, then review, then revise, and revision is heavy with local rearrangement. A document with thirty internal pastes and zero external pastes shows the rhythm of someone reading their own draft and moving pieces around. External pastes carry no such signal. They mean text from elsewhere was inserted here.

The two together — paste ratio AND paste source split — discriminate where either alone would be ambiguous. A document with a 6% paste ratio that's 100% internal looks like an editing-heavy session. The same 6% ratio at 100% external looks like material was imported from outside. Same bytes pasted; different stories. The composition fingerprint reads both.

Research history

Paste behavior as a discriminating signal has thinner published literature than keystroke dynamics or mouse geometry. The foundational HCI work treated copy-paste as an editing primitive to be improved, not a behavioral signature to be measured. Allen Cypher's Eager system (1991) catalogued paste-heavy editing patterns to predict and automate them; Robert Miller and Brad Myers's Outlier Finding work (2002) flagged unusual paste patterns as candidates for errors. Both were instrumentation studies, not authorship studies.

The cognitive-load grounding is older and better established. Miller's 1956 paper on the magical number seven gave the field a working-memory ceiling; the chunking literature that grew out of it — Simon (1974), Cowan (2001) — refined the estimate down to three to five items for unrehearsed material. Writing-process research (Hayes & Flower, 1980; Kellogg, 1996) mapped how composition actually unfolds in time, including the rearrangement-heavy revision phase that produces internal pastes.

Intrinsic plagiarism detection — the subfield concerned with detecting authorship inconsistencies inside a single document, rather than matching against external sources — uses paste-like signatures indirectly. Stein, Meyer zu Eissen, and Potthast's 2007 framing treats sudden stylistic shifts as evidence of imported content. The signal is text-based, not keystroke-based, but the underlying claim — that imported text reads differently from composed text — is the same. Shipping process-evidence tools agree in practice: Grammarly Authorship and Draftback both record and surface paste events as first-class provenance signals in their public documentation.

The field's state today: the mechanical claim that large external pastes are incompatible with in-flow composition is widely accepted on first-principles grounds. The empirical question — what threshold actually separates human-typical from model-typical sessions — has limited published calibration. Red Stet's sample corpus is one source of calibration data; we name it as such, openly.

Key papers

The paste-as-discriminator literature is thin. The works below either ground the mechanical claim (cognitive chunking, writing-process models) or document the signal in adjacent fields (plagiarism detection, HCI instrumentation). Real papers only.

Miller, G. A. (1956)

The magical number seven, plus or minus two: some limits on our capacity for processing information. Psychological Review, 63(2), 81–97.

Hayes, J. R., & Flower, L. S. (1980)

Identifying the organization of writing processes. In Cognitive Processes in Writing, Lawrence Erlbaum, 3–30.

Cypher, A. (1991)

Eager: programming repetitive tasks by example. Proceedings of CHI '91, 33–39. doi:10.1145/108844.108850

Cowan, N. (2001)

The magical number 4 in short-term memory: a reconsideration of mental storage capacity. Behavioral and Brain Sciences, 24(1), 87–114.

Miller, R. C., & Myers, B. A. (2002)

Outlier finding: focusing user attention on possible errors. Proceedings of UIST '02, 81–90. doi:10.1145/571985.571997

Stein, B., Meyer zu Eissen, S., & Potthast, M. (2007)

Strategies for retrieving plagiarized documents. Proceedings of SIGIR '07, 825–826.

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.

Red Stet calibration corpus (2026)

Internal calibration set — 11 documents: five public sample recordings in samples/ (four human, one generated under a synthetic profile), three onboarding training pieces, and three help-doc recordings. Used as one — not the only — source of threshold calibration. Acknowledged as a small, project-internal corpus; the full accounting is at Baselines.

Confidence level

Mixed: mechanical confidence high, empirical confidence lower.

The mechanical claim is hard to argue with. A four-hundred-character paste from the system clipboard, arriving in a single event, is incompatible with in-flow human composition no matter how the literature shakes out. Working memory does not produce four-hundred-character blocks. The writer either composed elsewhere and imported, or pulled the block from a non-compositional source. That's a physical-process argument, not a statistical one, and it holds without needing twenty papers behind it.

The empirical question — where exactly the threshold sits between "human-typical paste activity" and "uncommon for hand-typed work" — has less published calibration than keystroke cadence does. The numbers we use (paste ratio anchor at zero for full-human, 0.10 for fully machine-like; internal-share anchor at 100% for full-human, 0% for neutral) are calibrated mostly on the Red Stet sample corpus and the mechanical reasoning above. They will tighten as the corpus grows. We name that honestly here so the reader can weight the signal accordingly when reading the verifier's output.

The composite fingerprint does not lean on this signal alone. Paste ratio and paste source split together account for 27% of the composite weight — the largest combined block after keystroke cadence at 22%. That weighting reflects how loud the mechanical signal is, not how settled the empirical thresholds are. A flagged paste shows up in the right-panel timeline as a ●●● severity moment for the reader to inspect; the composite score is the summary, not the verdict.

Known limitations

The signal has real failure modes. The /science/ pages exist to name them, not to pretend they aren't there.

• Research writing legitimately pastes large blocks. Quoted passages, primary-source excerpts, code samples in a technical paper, block quotes in a literature review — all real human writing, all heavy with external pastes. Discrimination requires looking at the surrounding cadence and correction pattern, not the paste activity alone. A high paste ratio in a research session with normal typing cadence between the pastes reads differently from a high paste ratio with no typing in between.
• Translation work is paste-heavy by nature. A translator working through source text in another window will paste segments, type a translation, paste the next segment. The activity looks superficially like importing content; it isn't. We have no per-session way to distinguish "translation pass" from "imported draft" — the writer's own context is the disambiguator.
• Email and forum replies often start with a large paste. Quoting a prior message, pulling in a thread, pasting an error log — all common, all human-driven, all generate large external pastes that don't indicate model authorship.
• Adversaries can defeat this signal by typing model output character by character. Slow-typing pasted text removes the paste event from the record entirely. The composition fingerprint is the test, and the test runs on what the recording layer sees. We name this in the limitations page as the central adversary scenario, not an edge case.
• Internal vs clipboard isn't always cleanly distinguishable in the browser. The browser-side recorder does not flag this directly; instead, the verifier's analyzer derives the source category from cut/copy→paste proximity: a paste within 5 seconds of a recent cut or copy whose length matches within ±20% is classified as internal (copy-matching shipped 2026-06-11). The split is directional, not absolute — a writer who cuts a paragraph, pauses 6 seconds, then pastes it back will be misclassified as external, and a copy made in another window of similar length within the window would be misclassified as internal. The 5-second window keeps both error modes rare.
• Mobile keyboards insert chunks that look like pastes. Predictive-text completions, swipe input, voice-to-text dictation — all can produce multi-character inserts that the recording layer cannot reliably distinguish from a paste event without OS-level cooperation. The signal is calibrated on desktop sessions; mobile recordings need different anchors.
• The thresholds are calibrated on a small corpus. The Red Stet sample corpus is five human sessions and one model session at the time this page is written. Larger calibration sets — academic writing samples, professional editing sessions, ESL writer sessions — will move the anchors. We expect the threshold for "uncommon for hand-typed documents" to land lower (more permissive) as the corpus broadens.

How Red Stet uses it

The verifier scores both metrics on a 0–100 scale and folds them into the composite fingerprint with two separate weights.

Paste ratio score. Anchored at pasteRatio = 0 for the full human-typical score of 100 (no pasted characters at all) and at pasteRatio ≥ 0.10 for the machine-typical score of 0 (10% or more of the document arrived through the clipboard). The mapping is linear in between. Weight in the composite: 0.15.

Paste source score. The score is the internal share directly: 100% internal pastes scores 100; 0% internal — every paste came from outside the editor — scores 0. A session with no pastes at all scores a neutral 60: nothing to classify either way. Weight in the composite: 0.12.

Together the two account for 0.27 of the composite weight — the largest combined block after keystroke cadence. The weighting reflects the mechanical clarity of the signal: large external pastes are loud and hard to spoof without slow-typing, and slow-typing leaves its own traces in the cadence signal.

The right-panel timeline surfaces individual pastes as moments, not verdicts. A single paste of 480 characters from the clipboard at minute 1:14 appears in the Flagged column as a ●●● severity event with the timestamp, length, and source. The reader clicks through to the timeline marker and reads the surrounding context for themselves. The composite score summarizes; the timeline lets the reader verify. Per the 2026-06-08 framing decision, this surface produces evidence, not adjudication.