STORYPROOF — a keystroke-biometric CAPTCHA

Version: 0.1 · Status: prototype (static page, all scoring client-side) One line: Prove you're human by writing — a ~100-character story on a random easy prompt. It isn't what you write that proves humanity, it's how you type it, cross-checked against the fact that you actually answered this prompt.

---

1. Motivation

Traditional CAPTCHAs (image grids, distorted text) are increasingly solved by ML and sold as a service. Keystroke dynamics — the micro-timing of how a person presses and releases keys — is a behavioral biometric that is cheap to capture and hard to fake by hand. A free-text writing task is a natural vehicle: it produces a rich timing stream and content that can be sanity-checked.

But the research is unambiguous about a trap: timing alone is not sufficient. Multiple adversarial studies show that an attacker who samples inter-keystroke intervals from public human distributions defeats timing-only classifiers at ≥99.8% evasion, because keystroke timing carries zero mutual information about what was written — you can wrap any text (or replayed timing) in a human-looking rhythm.[B] So STORYPROOF deliberately combines three independent signal families and requires a session to look human on all three.

---

2. Threat model

Adversary	Technique	Primary defense
Naive script	el.value = "...", execCommand, dispatchEvent	Provenance: untrusted events, keystroke/char ratio
Paste / autofill	clipboard, password-manager, drag-drop	Provenance: paste/drop/insertFromPaste
Constant-delay bot	type with fixed setTimeout delay	Timing: variance (CV), repeated-interval detection
Fast bot	type faster than humanly possible	Timing: sub-50ms fraction, chars/sec ceiling
Statistical-mimicry bot	sample IKIs from human distributions[B][H]	Content: must answer this random prompt
Replay bot	replay a captured human's timing stream[I]	Content: replayed old text ≠ fresh random prompt
Headless / scripted mouse	jump cursor, straight-line moves, 0ms clicks	Mouse: curvature, velocity variance, teleports, click dwell
Form-filler	dutifully fills every field + rates + submits	Behavior: compliance probe — bots rate-then-verify; humans try to skip (§4.8)

The mimicry/replay rows are the dangerous ones, and the reason the random prompt + relevance check exists: a forged/replayed timing stream still has to emit text that responds to a prompt the attacker couldn't see in advance. That's the one thing timing forgery provably cannot supply.[B] The required rating adds a second adaptive step a naive form-filler won't satisfy (see §4.8).

---

3. What we capture (capture.js)

High-resolution, monotonic performance.now() timeline of every keyboard, mouse, and interaction event:

• keydown / keyup per key → dwell (hold = up − down) and flight / IKI (press-to-press gap, the standard inter-keystroke-interval measure).
• paste, drop, cut, compositionstart (IME), blur, focus.
• input with inputType (catches autofill/paste even when key events are absent).
• pointermove (throttled ~160 Hz, capped) → cursor path {t,x,y}; pointerdown / pointerup → click dwell and position; pointerType (mouse / touch / pen).
• slider input → rating value + change timeline (the required "rate this CAPTCHA").
• A unified interaction timeline of salient actions (focus, first keystroke, first mouse move, slider change, click) for measuring the gaps between interactions.
• isTrusted on every event (synthetic JS events are false) — incl. fake mouse moves.
• Counts: backspace, delete, paste, drop, untrusted, composition.

No data leaves the browser.

---

4. Empirical thresholds (and their sources)

The strongest numeric anchors are from Dhakal et al., CHI 2018 — the Aalto "136M Keystrokes" study of 168,000 participants, the largest keystroke corpus available.[A] Its task was transcription on physical/laptop keyboards, so absolute values transfer only approximately to free-text composition (composition yields more pausing and correction) and to mobile/touch keyboards — treat them as reference, not law.

4.1 Timing — human reference ranges (Dhakal CHI 2018 unless noted)

• Inter-key interval (IKI / flight): mean 238.66ms (SD 111.6), hard lower bound ~60ms; fast typists ~120ms (SD ~11), slow >480ms (SD >120). Distribution is right-skewed (skew ~2.0). Variance itself is the signal. STORYPROOF uses the coefficient of variation (CV=σ/μ): CV ≥ 0.5 solidly human, CV < 0.18 robotic.
• Dwell (key-hold): a tight motor constant — mean 116.24ms (SD 23.88), staying in the 80–150ms band even for slow typists. Dwell ≈ 0 or a single constant value ⇒ synthetic. The real tell is the within-band variance scripted input lacks.
• Digraph/bigram latency is structured, not uniform: hand-alternation 198ms, left-hand 215ms, right-hand 204ms, letter-repetition 176ms — key-pair- and hand-dependent. Uniform/random bot timings have no such structure.
• Sub-50ms gaps: ~5.82% of human intervals are <50ms vs ~21.49% for bots;[D] bots also hold keys briefly (23.11% of human keystrokes held <300ms vs 45.42% bot).[D] STORYPROOF flags if >25% of gaps are sub-50ms or sustained >22 chars/s (~216 wpm). (2013-era figures — evadable, but reliable against naive scripts.)

4.2 Rollover (key-overlap) — the strongest human-only signal

• Humans press the next key before releasing the previous one, so 25–50% of presses overlap (fast typists 40–70%; every typist group ≥19%).[A] Naive synthetic key injection produces 0% overlap. A practitioner scorer weights this highest (0.25 of its blend).[G] STORYPROOF detects it as the fraction of negative flight times and treats ≥20% as full credit, kept soft (slow hunt-and-peck and some mobile keyboards overlap less, so it lowers the score rather than hard-failing).

4.3 Cognitive pauses & bursts

• Natural composition is burst-and-pause: the keystroke-logging standard defines a burst as text with no IKI longer than 2 seconds; pauses >2s mark sentence/clause boundaries and index planning load, while average IKI stays stable across tasks.[E] Individualized thresholds (~2× the writer's median for sentence pauses, ~1.5× for within-word) beat a fixed cutoff.[E] For a short ~100-char story STORYPROOF uses a softer >500ms "micro-pause/planning" proxy and wants ≥1 — a 100-char passage with zero hesitations is suspiciously metronomic.

4.4 Correction behavior

• Correction is pervasive and a strong human signal: mean KSPC 1.173 (SD 0.094), ~6.3% of keystrokes corrected, 2.29 corrections/sentence (99th pct ~8.5 backspaces/sentence).[A] Near-zero backspaces or KSPC = 1.0 is atypical of genuine typing (and composition corrects more than the transcription task these come from). A true "revision" is deletion plus retyped replacement, not a lone backspace.[F] STORYPROOF treats corrections as a soft positive (absence is a warning, not a hard fail — a clean short sentence is plausible).

4.5 Why content is mandatory, not optional

• Timing alone is evadable: statistical forgeries reach ~15% false-acceptance when the attacker has target-user data, and a dedicated liveness classifier hits only 1–2% FAR/FRR against weak (general-population-data) attackers.[C] Because keystroke timing carries zero information about what was written, the random prompt + relevance check is the cheap, backend-free defense forgery/replay cannot satisfy — it must produce text answering a prompt the attacker couldn't see in advance.

4.6 Mouse dynamics

• The behavioral-biometrics survey literature defines the discriminating features straight from raw (x,y,t):[J] curvature κ = 1/r (exactly 0 for a straight line), straightness / efficiency = displacement ÷ path-distance (→ 1.0 for bot-like straight paths), tremor/jitter = interpolated ÷ actual path length (→ 1.0 = no micro-jitter), plus velocity, acceleration, tangential jerk, angular velocity distributions. Human motion also has neuromotor structure (Sigma-Lognormal): an initial acceleration burst and a final low-velocity fine-correction as the cursor nears the target — "difficult to emulate for bots"; a single trajectory scores ~93% human-vs-bot in BeCAPTCHA-Mouse.[K]
• Clicks are their own signal: mousedown→mouseup dwell, time-to-click, pause count, and paused-ratio.[J] Bots show no hover-before-click and ≈ 0ms dwell.
• STORYPROOF computes a usable subset only when a real mouse/pen moved: curvature (mean direction change/sample + straightness), velocity-CV, teleports (>160px jump in <16ms = no intermediate samples), and click dwell.
• ⚠️ No hard velocity cutoffs. Verification refuted the specific 2013 numbers (human ~427 px/s vs bot ~1521 px/s; efficiency > 0.94 cutoff) — treat speed/efficiency as directional, not authoritative. STORYPROOF keys on variance and structure, not absolute px/s.
• Caveat: touch and keyboard-only users produce no usable cursor path — those sessions mark the mouse signal n/a and reweight onto keystrokes (never a hard fail).

4.7 Interaction timing

• Distinct UI actions are separated by human reaction time: mean simple reaction time is ~210–240ms with a hard anticipatory floor near 110ms — responses faster than ~110ms are excluded as non-genuine.[L] STORYPROOF builds a unified timeline (focus → first key → mouse → slider → click) and flags sessions where a large fraction of inter-action gaps fall below ~120ms (sub-reaction-time) or have near-zero spread — the signature of a script firing actions instantly or at one uniform delay.

4.8 Behavior — the required rating as a compliance probe

• A 1–10 rating slider defaults to 10 and is required to complete — but the page never says so until "Verify I'm human" is pressed. The discriminator is not whether you rate it (everyone must, eventually) — it's whether you tried to skip it:
  • A bot is dutiful: it sees the rating control, rates it, then clicks verify. It never trips the hidden requirement → scored as compliant (bot-leaning, −5).
  • A human dodges busywork: they hit "Verify I'm human" and ignore the rating; the page then reveals it's required and forces them to rate to continue. That skip attempt (verify-before-rate) is the human tell → +8.
• It is a weak, deliberately playful signal — it nudges the score, adds an adaptive step and extra interaction-timing surface; it is not load-bearing, and a bot that learns to verify-first defeats it. (Reaction-time of the skip→forced-rate sequence is also captured by §4.7.)

---

5. Scoring (score.js)

Six families. Provenance and content are gates; timing, mouse, and interaction are blended; behavior nudges.

A. Provenance (hard gates — any one ⇒ reject):

1. No paste / drop / insertFromPaste-style input.
2. No untrusted (synthetic) events — incl. fake pointermove.
3. keystrokes/char ratio ≥ 0.5 (else value was set programmatically).
4. ≥70 chars and ≥40 keystrokes (else "not enough" — forgiven only as data, not a tell).

B. Keystroke timing (blended, weights below):

Signal	Weight	Human-pass	Bot-fail
Rhythm variability (CV)	0.22	CV ≥ 0.5	CV < 0.18
Plausible speed	0.15	<12 chars/s, <8% sub-50ms	>22 chars/s or >25% sub-50ms
Dwell profile	0.16	median 25–250ms, σ > 5ms	dwell ≈ 0 / constant
Key-overlap (rollover)	0.15	≥20% negative flights	0% overlap
Non-repeating intervals	0.14	>50% unique gap values	<50% unique (metronomic)
Cognitive pauses	0.10	≥1 pause >500ms	none
Correction behavior	0.08	≥1 backspace/delete	none (soft)

C. Content (anti-replay / anti-mimicry):

1. Reads like language — ≥8 words, avg word length 2.5–12, unique-word ratio >0.35.
2. On-topic — overlaps ≥1 content word with the randomized prompt.

D. Mouse (0..1, only if a real mouse/pen moved): curvature 0.4 · velocity-CV 0.3 · no-teleport 0.2 · click-dwell 0.1. E. Interaction timing (0..1): (1 − sub-120ms fraction) · 0.6 + (gap-CV) · 0.4. F. Behavior: +8 if they tried to skip the rating (hit verify before rating — human), −5 if they rated then verified (compliant — bot).

Composite: base = mouseUsed ? ks·0.55 + mouse·0.25 + interaction·0.20 : ks·0.75 + interaction·0.25, then ×100 + slider bonus. Floors: !sane → ≤20, tooFast || metronomic → ≤35, any hard provenance fail → ≤8; minus a relevance penalty.

Verdict bands: human ≥ 70 · review 48–69 · bot < 48. (review = borderline; a real deployment escalates to a second challenge.)

---

6. Known limitations (be honest)

• CDP/driver input can be trusted. Playwright/Puppeteer type() produces trusted events and can inject per-key delay jitter. STORYPROOF catches naive versions via variance + content, but a tool that samples human IKIs and writes a relevant story on the fly would pass — consistent with the ≥99.8% timing-evasion result.[B] The random prompt raises the cost (the attacker needs real-time relevant generation), it doesn't make it impossible. A production system pairs this with rate limiting, proof-of-work, and server-side replay/nonce checks.
• Accessibility. Switch access, dictation, on-screen keyboards, and IME users produce atypical timing. The review band and a non-keystroke fallback challenge are required for a real deployment — never hard-fail on timing alone.
• Mobile / touch / trackpad. Touch keyboards have different dwell/flight profiles and autocorrect rewrites; touch produces no cursor path (mouse signal goes n/a); trackpad motion is choppier than a mouse. Thresholds here are tuned for physical keyboard + mouse and would need device-class profiles.
• Mouse mimicry. Humanlike-cursor libraries (e.g. ghost-cursor) and GAN/RNN trajectory synthesis reproduce curvature, tremor, and velocity profiles — so mouse dynamics, like keystroke timing, defeat naive bots but not a determined mimic. It is one weak signal in the blend, never a gate.
• The required-rating compliance probe is theater, not security. It exploits that naive bots dutifully rate-then-verify while humans try to skip the busywork — but a bot that learns to click verify first defeats it, and an over-eager human who rates first gets dinged (−5, never fatal). Treat it as a tarpit/telemetry nudge, not a real barrier.
• Client-side scoring is inspectable. This prototype scores in-browser for demo transparency (window.__storyproof). Production must score server-side from the raw event log so the attacker can't read the rubric.

---

7. Files

File	Role
index.html	Page shell + prompt UI
styles.css	Cyber-brutalist dark theme
topics.js	Random easy-prompt bank + picker
capture.js	Keystroke + mouse + interaction-timeline + slider recorder (capture only)
score.js	Six-family humanness scorer + thresholds
app.js	Wiring: prompt → record → score → verdict; required-rating gate
test-score.mjs	Node fixture tests (keystroke / mouse / interaction / slider)

---

8. Sources

Synthesized from two fan-out research passes — keystroke dynamics (19 sources, 88 claims, 25 verified → 20 confirmed) and mouse/interaction dynamics (20 sources, 87 claims, 25 verified → 21 confirmed). Confirmed claims drive the numbers above; the refuted ones are why §4 avoids brittle single cutoffs.

Keystroke dynamics [A–I]:

• [A] Dhakal, Feit, Kristensson & Oulasvirta, Observations on Typing from 136 Million Keystrokes, CHI 2018 (n=168,000). The primary anchor — IKI 238.66ms (SD 111.6), dwell 116.24ms (SD 23.88), hand-dependent bigram latencies, rollover 25–70%, KSPC 1.173. dl.acm.org/doi/pdf/10.1145/3173574.3174220
• [B] On the Insecurity of Keystroke-Based AI Authorship Detection: Timing-Forgery Attacks — arXiv:2601.17280. Three forgery attacks reach ≥99.8% evasion of five classifiers; timing carries zero content information.
• [C] Gonzalez et al., Spoofing & liveness in free-text keystroke dynamics, Systems and Soft Computing 4 (2022), Elsevier. Best forgery ~15% FAR with target-user data; liveness classifier 1–2% FAR/FRR only against weak attackers. sciencedirect.com/science/article/pii/S2772941922000047
• [D] Chu, Wang et al., Blog-bot detection via keystroke/timing, Computer Networks 2013, Elsevier. Sub-50ms gaps 5.82% human vs 21.49% bot; dwell <300ms 23.11% human vs 45.42% bot; fixed-timer clustering ⇒ low entropy. eecis.udel.edu/~hnw/paper/comnet13.pdf
• [E] Van Waes et al., Bursts & pauses in writing (Reading and Writing 2019, Springer) + Rosenqvist 2015 — 2s burst boundary; individualized pause thresholds (~2× median sentence, ~1.5× within-word). link.springer.com/article/10.1007/s11145-019-09953-8
• [F] Conijn et al., Automated extraction of revision events from keystroke data (Reading and Writing 2021, Springer) — a revision = deletion + retyped replacement, not a lone backspace.
• [G] isHumanCadence (open-source) — practical six-signal cadence scorer that weights rollover highest (0.25) and implements paste/constant-interval/uniform-jitter/replay detectors. Non-peer-reviewed; its weights are design choices. github.com/RoloBits/isHumanCadence
• [H] Conditional GAN for Keystroke Presentation Attack (arXiv:2212.08445) + Stefan, Shu & Yao, Robustness of Keystroke-Dynamics Against Synthetic Forgeries (GaussianBot / NoiseBot) — statistical/GAN synthesis of human-like dwell/flight.
• [I] Spoofing keystroke dynamics … from screen-recorded video (J. Big Data 2022, Springer) — replay of genuine captured timing.

Mouse & interaction dynamics [J–M]:

• [J] Khan, Devlen, Manno & Hou, Mouse Dynamics Behavioral Biometrics: A Survey, ACM Computing Surveys 2024 (DOI 10.1145/3640311 = arXiv:2208.09061). The feature taxonomy — curvature κ=1/r, straightness=displacement/path-distance, tremor=interpolated/ actual path, velocity/acceleration/jerk/angular-velocity, and click dwell/pause features.
• [K] Acien, Morales, Fierrez & Vera-Rodriguez, BeCAPTCHA-Mouse, Pattern Recognition 2022 (arXiv:2005.00890). Sigma-Lognormal neuromotor model — human initial-acceleration + final low-velocity fine-correction; ~93% single-trajectory bot detection (own benchmark; velocity-profile synthetics still fooled it up to ~17%).
• [L] Woods et al., Factors influencing simple reaction time, Front. Hum. Neurosci. 2015 (PMC4374455, n=1,469). Mean SRT ~231ms (213 hardware-corrected); 110–1000ms valid response window — the ~110ms anticipatory floor grounds the sub-reaction-time heuristic.
• [M] ghost-cursor (github.com/Xetera/ghost-cursor) + DMTG: diffusion mouse-trajectory generator (arXiv:2410.18233, 2024). The mimicry frontier — Bézier + Fitts's-law point counts + deliberate overshoot/re-adjust; diffusion/GAN synthesis degrades detectors 4.75–9.73%. Why mouse dynamics is a weak blended signal, never a gate.

Verification refuted (do not implement): (keystroke) a fixed flight-SD <20ms cutoff; a per-keystroke "sub-60ms is physically impossible" rule; "100% detection" claims; and "zero corrections is normal for half of typists" (it is atypical). (mouse) the 2013 absolute velocity cutoffs (human ~427 vs bot ~1521 px/s, ~3500 px/s cap) and the straightness=1.0 / efficiency>0.94 hard thresholds — directional only. All are why STORYPROOF scores aggregate distributions and blends many weak signals instead of trusting one hard line.