Investigation thresholds and baseline rates

The flag-rate framing

A flagged event in a Red Stet record is a single moment — one keystroke run with collapsed cadence, one paste of clipboard text, one mouse path that lacks the curvature of a hand-driven trackpad. A raw count of those moments is uninterpretable on its own. A 5,000-word draft produces on the order of 30,000 composition events (keystrokes, deletions, pastes, cuts); two flagged moments in it is under 0.01% of the writing activity — statistical noise. The same two flags in a 150-word note, perhaps 1,000 composition events, is a different kind of object. The threshold has to live on the rate, not the raw count.

The right denominator is composition events — keystrokes, deletions, pastes, and cuts — not document length, not flagged-event count, and not total recorded events. Every threshold on this page is expressed as a percentage of composition events. This denominator is load-bearing for the rest of the methodology: it normalizes across document length, writing speed, and the natural variance in event density between a haiku draft and a long-form essay — and, critically, it does not move when cursor-sampling fidelity changes. The recorder samples the cursor at roughly 60 Hz; counting those samples in the denominator would deflate every flag rate as mouse activity rose, so a fidgety session would read "cleaner" than a still one with identical text behavior. Composition events sidestep that entirely.

The verifier's right panel reflects this. The three sections — Flagged, Questionable, Key Human — show both the count and the rate against composition events, computed over all detected moments (the timeline display caps how many moments render, but the rates are computed before that cap). One precision note: the composite verdict tier is computed from the seven weighted signal scores, not from the flag rate — the rate and the composite are two independent views of the same recording, and this page's threshold ladder is reviewer guidance for reading the rate, not a second input to the tier.

What research provides directly

The behavioral-biometrics literature is rich on within-population distributions. Forty-five years of work — Gaines and colleagues in 1980, Joyce and Gupta in 1990, Monrose and Rubin through the late 1990s, Gunetti and Picardi in the 2000s, Killourhy and Maxion's 2009 benchmark — has measured what hand-typed English text looks like across thousands of subjects. Keystroke inter-key σ between 30 and 80 ms. Backspace rates between 1.5% and 3% of keystrokes. Paste activity dominated by single-token snippets, not paragraph-scale blocks. Each per-signal page in this methodology section cites that literature in detail.

What the field does not provide is a clean between-population threshold for human-versus-machine text generation. The published corpora were assembled before large language models became the adversary the field now thinks about. Killourhy-Maxion's CMU dataset is fixed-string login authentication. Gunetti-Picardi's free-text corpus is human-to-human identification, not human-to-machine. There is recent and growing work on detecting model output from its textual surface — but that work is on the prose, not on the process events Red Stet records. The bridge from "this is the within-human distribution" to "this is the threshold above which we should investigate further" has to be built by aggregating the per-signal distributions and calibrating a composite against ground-truth records.

Red Stet's per-signal pages do the first half of that work: each page synthesizes the published within-human distribution for one signal and names the score anchors derived from it. This page does the second half: it assembles those signals into a composite, calibrates the thresholds against a small ground-truth corpus, and is honest about which numbers are read from the published literature and which are read from our corpus.

The calibration corpus

Red Stet's calibration corpus is eleven documents, in three groups. Five public sample documents in samples/: four human-authored — a YA novel chapter (Caleb Morrow, 2,504 words), a literary short fiction (Daniel Reisman, 948 words), an op-ed by a small-town board chair (Eleanor Hendricks, 1,303 words), and a haiku cycle with prose interludes (Margaret Wheeler, 415 words) — plus one narrative-nonfiction chapter (K. Anders, 1,706 words) produced under an explicit synthetic profile: zero backspaces, one large clipboard paste at 32% through, two-point mouse paths, ~45 ms keystroke cadence. Three short training pieces written through Red Stet's onboarding flow. Three representative help documents (first-class, peer-review, writers-first-doc) whose provenance records were generated through the same pipeline so the help corpus has matched ground-truth structure. Where other methodology pages mention "the sample corpus," they mean the five public documents; the thresholds on this page are calibrated against all eleven. The folder layout, the manifest, and the per-document statistics are public at github.com/Moikeha09/redstet/tree/main/samples.

Ten of the eleven documents are human-authored. One is machine-generated under a deterministic profile that emits the patterns published research describes as machine-typical. The per-signal score anchors in the verifier are set just outside the corpus clusters: the σ = 60 ms full-credit anchor sits below the human samples (which cluster at σ ≈ 89–97 ms in the public recordings, recomputed 2026-06-10); the σ = 8 ms zero anchor sits below the synthetic sample (σ ≈ 10 ms). The backspace anchor of 15/kchar scoring 100 sits at the lower edge of human samples that cluster at 19–22/kchar. Every anchor on every per-signal page traces back to this corpus.

The corpus is small, English-language, and adversarial-poor. There is no sample produced by a human who read model output and re-typed it at human cadence — the canonical adversary against affirmative-human framing. The corpus is heavy on literary prose and lighter on academic, technical, and journalistic writing styles. It does not include voice-to-text dictation, swipe-typing, or non-Roman-orthography sessions. These are real gaps. The right response is not to overclaim what the corpus supports — it is to name the gaps explicitly here and at /science/limitations/, ship thresholds calibrated against what we have, and tighten them as the corpus grows.

Eleven documents is enough to set anchors that separate cleanly in the corpus — the human documents score 81–96 on the composite; the synthetic document scores 24. The gap is real and reproducible (the analyzer is deterministic; the five public samples can be re-run through the verifier or the samples/ CLI by anyone). Eleven documents is not enough to claim a calibrated false-positive rate against the worldwide distribution of writing styles. The thresholds below are descriptive of where our corpus sits; they are not population-calibrated.

The investigation thresholds

An investigation threshold is the flag rate above which a document is worth examining in detail and below which the flag count is statistical noise within the natural variance of hand-typed writing. Red Stet uses four bands, expressed as the share of recorded events the verifier flagged in the document.

The bands are descriptive of the calibration corpus. The 1% lower edge is where our ten human documents sit comfortably — the median human document in our corpus produces flagged-event rates well below 1% of the total stream, even when the verifier surfaces individual moments worth a brief look. The 5% upper edge is where our single AI sample sits and above. The 2% mid-edge — the threshold this page exists to name — is where the published per-signal distributions stop producing flags in the natural course of human typing and start indicating something structurally different about the session.

The flag rate and the composite score are independent numbers, and the right reviewer move is to read them together. A document at 3% flag rate with composite 75 deserves a review of its flagged moments even though the overall pattern reads human; a document at 3% with composite 35 points the same direction twice. To be precise about mechanism: the system does not combine the two — the composite tier comes from the weighted signals alone, and these bands are published guidance for the reviewer's reading of the rate. When the two views disagree, the moments timeline is where the disagreement gets resolved.

Two notes on what these thresholds are not. They are not population false-positive rates — we do not claim that a document below 1% has a 1% chance of being model-generated. They are not prescriptive standards an integrity board should adopt as policy — that question requires more corpus, more adversarial samples, and a deployment phase the product has not reached. They are descriptive markers calibrated against eleven documents, published openly so the reader can audit them, and tightened as the calibration corpus grows.

Composite-score thresholds

The composite is a weighted average of seven per-signal scores, each on a 0–100 scale. The weights — keystroke cadence 0.22, backspace rate 0.20, paste ratio 0.15, paste source 0.12, thinking pauses 0.11, mouse path geometry 0.10, click position 0.10 — total to 1.00 and are visible in the analyzeVoice function of src/provenance/analysis.mjs. Cadence and backspaces carry the most weight because their underlying physiology is the hardest to spoof at scale and the published literature on them is the deepest. The other five signals are corroborative.

Three composite tiers map to the verdict language used everywhere else in the product:

The 70 and 40 cutoffs sit in the empty range between the corpus clusters. The human documents land between 81 and 96; the synthetic document lands at 24. Eighty-one to seventy is an 11-point buffer; forty to twenty-four is a 16-point buffer. As the calibration corpus grows — adversarial samples, cross-language samples, voice-to-text sessions — these cutoffs will shift to track the empirical distribution. The current placement is conservative on both sides: a document has to land well into the human range to read as "consistent" and well into the machine-like range to read as "uncommon."

The composite score and the flag rate are two views of the same evidence. The composite weights the signals by published reliability; the flag rate counts moments without weighting them. When the two agree, the verdict is clean. When they disagree — a high composite with elevated flag rate, or a low composite with a low flag rate — the verifier's right-panel timeline is where the reader resolves the case. The numbers narrow the question; they do not adjudicate it.

How these thresholds should evolve

Denominator fix landed (2026-06-15). The flag-rate bands were originally expressed as a share of total recorded events, a denominator sensitive to recorder fidelity: when the cursor sample rate rose from ~4 Hz to ~60 Hz in June 2026, the cursor-event share of a typical recording jumped and deflated flag percentages. The rate is now computed against composition events (keystrokes, deletions, pastes, cuts) — a denominator that does not move when cursor fidelity changes. Re-running the analyzer over the eleven-document corpus under the new denominator left the bands intact: the human documents still sit well below 1% and the synthetic profile above 5%, because those recordings carry little cursor activity, so total and composition events nearly coincide for them. The change matters for real, mouse-heavy sessions going forward, where the old denominator understated the rate. The band edges below are unchanged.

Beyond that, real-world deployment generates new data. The thresholds on this page should tighten in four directions as the calibration corpus grows.

Adversarial corpus expansion. The canonical adversary against affirmative-human framing is a person who reads model output and re-types it at human cadence. The published literature is mostly silent on this case because it predates the relevant threat. Red Stet's corpus does not include such samples either. Adding them is the highest-priority calibration work: it will either confirm the composite distinguishes the case via paste source, correction rate, and thinking-pause structure, or it will name an attack the composite does not catch. Either result is more useful than the current state of unknown.

Cross-language calibration. The Bergadano-Gunetti-Picardi line of work showed the keystroke-cadence signal survives Italian. Equivalent calibration for the other six signals across Spanish, Mandarin pinyin input, Japanese kana-to-kanji conversion, and right-to-left scripts has not been done in Red Stet's corpus. Backspace rates in particular vary across input methods in ways the current thresholds do not account for.

Domain-specific calibration. Academic writing produces different baseline distributions than fiction: more pasted citations, longer thinking pauses, more focus-switching to external sources. Technical writing produces still others: more terminology runs, more code blocks, different correction patterns around symbol-heavy passages. A unified threshold across all domains over-flags the styles that diverge from the corpus median. Domain-specific anchor sets are the right correction.

Anonymous opt-in writer data. Writers who consent to contribute their compositions to the calibration corpus would let Red Stet measure the empirical distribution across a population larger than eleven documents. This is a privacy-careful path — the data shape required is summary statistics per session, not raw event streams — and it is the only way to move from descriptive thresholds against a small corpus to population-calibrated thresholds against a large one. Until that path is built, the thresholds on this page remain explicitly descriptive.

Bottom line

Investigation begins at flag rates above 2% of recorded events for human-authored documents. Below that threshold, flagged events are statistical noise within the natural distribution of behavior in hand-typed writing. The composite score thresholds — 70 and 40 — narrow the verdict between the three published tiers. These thresholds are descriptive of Red Stet's current eleven-document calibration corpus, not prescriptive standards derived from independent population research. The per-signal pages cite the published literature behind each underlying distribution; this page names the calibration step that builds composite thresholds on top of them. The right reading is: the signals exist in the literature, the anchors are read from our corpus, the thresholds are honest about which is which, and the verifier surfaces enough moment-level evidence for a reader to form an independent judgment on any document the rate-and-composite combination does not resolve cleanly.