Cognitive Typing Rules

• Cognitive Typing Rules are formal methods that link keystroke timing with content complexity to reveal cognitive processes like planning, translating, and revising.
• They employ metrics such as the Cognitive Load Correlation (CLC) to distinguish genuine composition from mechanical transcription based on timing-content covariance.
• These rules integrate preprocessing, phase segmentation, and quantization to ensure privacy and robust non-intrusive authorship verification in writing interfaces.

Searching arXiv for the cited papers to ground the article. “Cognitive typing rules” denotes a family of formal rules that connect observable sequences—most prominently keystroke timing, but also event sequences, conceptual transformations, and typographic design variables—to latent cognitive structure. In the keystroke-authentication formulation, the term refers to rules for inferring planning, translating, and revising behavior from inter-keystroke intervals and then using their coupling to local content complexity for non-intrusive authorship verification (Condrey, 26 Feb 2026). In related arXiv work, the same phrase or closely related usage appears in cognitively motivated event-process typing, type-theoretic models of scientific concept formation, computational models of touchscreen typing errors, and computational typography, indicating a broader methodological theme: cognition is modeled through typed relations between observable structure and higher-order intent, effort, or constraint (Chen et al., 2020).

1. Definition and Scope

In “Detecting Cognitive Signatures in Typing Behavior for Non-Intrusive Authorship Verification” (Condrey, 26 Feb 2026), cognitive typing rules are defined around the claim that ordinary typing interfaces capture “rich cognitive signatures,” measurable as patterns in keystroke timing that reflect the planning, translating, and revising stages of genuine composition. The central object is the Cognitive Load Correlation (CLC), a Pearson correlation between inter-keystroke intervals and local content-complexity scores. Under this formulation, authorship verification is reframed from output analysis to human-computer interaction.

The same expression has a broader cross-domain presence. In “What Are You Trying to Do?” (Chen et al., 2020), a cognitively motivated typing task assigns free-form action and object labels to event processes by inferring goals and affected object types. In “How are Scientific Concepts Birthed?” (Aguilar et al., 12 Sep 2025), cognitive typing rules are type-theoretic inference rules for concept distinction, property preservation, and concept change. In “Simulating Errors in Touchscreen Typing” (Shi et al., 5 Feb 2025), the phrase is used for rules governing slips, lapses, and mistakes under supervisory control. A plausible implication is that “cognitive typing rules” is best understood not as a single standardized term, but as a recurring formalization strategy in which typed observables are used to recover latent cognitive organization.

Within the keystroke-authorship setting, the rules are explicitly non-intrusive. The framework operates within existing writing interfaces, collects timing metadata only, and is designed to preserve privacy through quantization and aggressive data minimization (Condrey, 26 Feb 2026).

2. Core Formalism: Cognitive Load Correlation

The foundational quantity is the Cognitive Load Correlation, denoted $\mathrm{CLC}$, defined as the Pearson correlation coefficient between two per-keystroke sequences: $t_i$, the measured inter-keystroke interval at keystroke index $i$, and $c_i$, the estimated content-complexity at that same position. Formally,

$$\mathrm{CLC} \;=\; \rho_{t,c} \;=\; \frac {\displaystyle \sum_{i=1}^N \bigl(t_i - \bar t\bigr)\,\bigl(c_i - \bar c\bigr)} {\sqrt{\displaystyle\sum_{i=1}^N (t_i - \bar t)^2} \,\sqrt{\displaystyle\sum_{i=1}^N(c_i - \bar c)^2} }\,.$$

Here, $N$ is the total number of keystrokes in the document; $t_i$ is the inter-keystroke interval in milliseconds between key $i-1$ and key $i$; $c_i$ is a scalar content-complexity score at position $t_i$0; $t_i$1 is the mean IKI; and $t_i$2 is the mean complexity (Condrey, 26 Feb 2026).

The interpretation supplied for this statistic is operational rather than merely descriptive. Genuine composition yields moderate-to-strong positive CLC, typically $t_i$3, because the writer pauses longer before more complex passages. Mechanical transcription yields near-zero correlation, $t_i$4, because timing is governed by motor fluency rather than content generation. This is the principal distinction encoded by the rules: composition is treated as a coupled cognitive-semantic process, whereas transcription is treated as a largely motor sequence.

The content-complexity signal $t_i$5 may be instantiated by an n-gram surprisal or a syntactic depth score. This means the framework is not confined to one linguistic theory of difficulty; instead, it requires only a per-position scalar estimate that can serve as a proxy for local content difficulty. This suggests that the rules are modular with respect to the language-modeling component, while remaining fixed in their dependence on timing-content covariance.

3. Keystroke Preprocessing and Cognitive Phase Segmentation

Before CLC is computed, the raw keystroke stream is preprocessed and segmented into three cognitive phases—planning, translating, and revising—using timing and key-event metadata alone (Condrey, 26 Feb 2026). The raw input is a timestamped event stream

$t_i$6

with millisecond resolution. Inter-keystroke intervals are then extracted from successive key-down times $t_i$7 as

$t_i$8

Planning is identified by long pauses. A keystroke $t_i$9 is marked as a planning boundary if

$i$0

with $i$1 and configurable between $i$2 and $i$3 seconds. Translating is defined at the burst level. Between planning boundaries, contiguous keystrokes form a burst; when the median IKI inside the burst satisfies

$i$4

with $i$5, that region is labeled TRANSLATING. Revising is detected from deletion behavior: any interval containing a sequence of deletion events of length at least $i$6 within a short window $i$7 is labeled REVISING, with default parameters $i$8 deletions in $i$9 ms.

The associated high-level procedure returns not only $c_i$0 but also phase labels and revision density, where $c_i$1. Unknown regions are assigned TRANSLATING by default. The resulting segmentation is used both for visualization and for higher-order feature extraction. In this rule system, cognitive phases are therefore not latent states inferred by a deep sequential model; they are explicit labels induced by thresholded temporal structure and event metadata.

4. Decision Rules for Authorship Verification

Once $c_i$2 is computed, the framework recommends a simple decision rule (Condrey, 26 Feb 2026). If $c_i$3, the session is classified as Genuine Composition. If $c_i$4, it is classified as Mechanical Transcription. Values in $c_i$5 are treated as borderline and are said to warrant further evidence, such as revision density or session-consistency checks.

These thresholds are tied to analytical distributions given in the paper: genuine composition has $c_i$6 with SD $c_i$7, while transcription has $c_i$8 with SD $c_i$9. The paper synthesizes published findings to derive an estimated discrimination accuracy of 85–95 percent AUC. The lower bound is associated with population heterogeneity, including novice typists and L2 writers, whereas the upper bound is associated with controlled laboratory studies using experienced typists and well-annotated content. For documents of at least 1,500 words, statistical power exceeds $$\mathrm{CLC} \;=\; \rho_{t,c} \;=\; \frac {\displaystyle \sum_{i=1}^N \bigl(t_i - \bar t\bigr)\,\bigl(c_i - \bar c\bigr)} {\sqrt{\displaystyle\sum_{i=1}^N (t_i - \bar t)^2} \,\sqrt{\displaystyle\sum_{i=1}^N(c_i - \bar c)^2} }\,.$$0 under the stated assumption $$\mathrm{CLC} \;=\; \rho_{t,c} \;=\; \frac {\displaystyle \sum_{i=1}^N \bigl(t_i - \bar t\bigr)\,\bigl(c_i - \bar c\bigr)} {\sqrt{\displaystyle\sum_{i=1}^N (t_i - \bar t)^2} \,\sqrt{\displaystyle\sum_{i=1}^N(c_i - \bar c)^2} }\,.$$1.

The statistical significance argument is similarly explicit. Under a two-sample t-test on $$\mathrm{CLC} \;=\; \rho_{t,c} \;=\; \frac {\displaystyle \sum_{i=1}^N \bigl(t_i - \bar t\bigr)\,\bigl(c_i - \bar c\bigr)} {\sqrt{\displaystyle\sum_{i=1}^N (t_i - \bar t)^2} \,\sqrt{\displaystyle\sum_{i=1}^N(c_i - \bar c)^2} }\,.$$2 values from the two conditions, a difference of means of approximately $$\mathrm{CLC} \;=\; \rho_{t,c} \;=\; \frac {\displaystyle \sum_{i=1}^N \bigl(t_i - \bar t\bigr)\,\bigl(c_i - \bar c\bigr)} {\sqrt{\displaystyle\sum_{i=1}^N (t_i - \bar t)^2} \,\sqrt{\displaystyle\sum_{i=1}^N(c_i - \bar c)^2} }\,.$$3 with pooled SD approximately $$\mathrm{CLC} \;=\; \rho_{t,c} \;=\; \frac {\displaystyle \sum_{i=1}^N \bigl(t_i - \bar t\bigr)\,\bigl(c_i - \bar c\bigr)} {\sqrt{\displaystyle\sum_{i=1}^N (t_i - \bar t)^2} \,\sqrt{\displaystyle\sum_{i=1}^N(c_i - \bar c)^2} }\,.$$4 yields $$\mathrm{CLC} \;=\; \rho_{t,c} \;=\; \frac {\displaystyle \sum_{i=1}^N \bigl(t_i - \bar t\bigr)\,\bigl(c_i - \bar c\bigr)} {\sqrt{\displaystyle\sum_{i=1}^N (t_i - \bar t)^2} \,\sqrt{\displaystyle\sum_{i=1}^N(c_i - \bar c)^2} }\,.$$5 for $$\mathrm{CLC} \;=\; \rho_{t,c} \;=\; \frac {\displaystyle \sum_{i=1}^N \bigl(t_i - \bar t\bigr)\,\bigl(c_i - \bar c\bigr)} {\sqrt{\displaystyle\sum_{i=1}^N (t_i - \bar t)^2} \,\sqrt{\displaystyle\sum_{i=1}^N(c_i - \bar c)^2} }\,.$$6 per group, with $$\mathrm{CLC} \;=\; \rho_{t,c} \;=\; \frac {\displaystyle \sum_{i=1}^N \bigl(t_i - \bar t\bigr)\,\bigl(c_i - \bar c\bigr)} {\sqrt{\displaystyle\sum_{i=1}^N (t_i - \bar t)^2} \,\sqrt{\displaystyle\sum_{i=1}^N(c_i - \bar c)^2} }\,.$$7. The framework therefore treats the distinction not as a stylistic signal of final text, but as a behavioral signature in the generation process itself.

5. Privacy Preservation and Adversarial Robustness

A defining feature of the keystroke-based rules is privacy preservation through evidence quantization. Before any further processing, inter-keystroke intervals are quantized by

$$\mathrm{CLC} \;=\; \rho_{t,c} \;=\; \frac {\displaystyle \sum_{i=1}^N \bigl(t_i - \bar t\bigr)\,\bigl(c_i - \bar c\bigr)} {\sqrt{\displaystyle\sum_{i=1}^N (t_i - \bar t)^2} \,\sqrt{\displaystyle\sum_{i=1}^N(c_i - \bar c)^2} }\,.$$8

At $$\mathrm{CLC} \;=\; \rho_{t,c} \;=\; \frac {\displaystyle \sum_{i=1}^N \bigl(t_i - \bar t\bigr)\,\bigl(c_i - \bar c\bigr)} {\sqrt{\displaystyle\sum_{i=1}^N (t_i - \bar t)^2} \,\sqrt{\displaystyle\sum_{i=1}^N(c_i - \bar c)^2} }\,.$$9 ms, the analysis cited from Killourhy and Maxion (2008) is reported to show that motor-based verification virtually collapses, with EER approaching approximately $N$0, while cognitive pauses of at least $N$1 ms are preserved with less than $N$2 distortion (Condrey, 26 Feb 2026). The rule system is therefore explicitly designed to suppress motor-level biometric leakage while retaining slower, cognitively meaningful timing structure.

The adversarial robustness argument proceeds in several layers. First, motor-only timing is described as low-dimensional and forgeable; prior work is cited as showing greater than $N$3 evasion of identity authentication using purely motor-level forgery. Second, cognitive timing is presented as multilevel and content-dependent: pauses must correlate with lexical surprisal, burst lengths must align with clause boundaries, and revision patterns must mimic genuine real-time evaluation. Third, the asymmetry is expressed information-theoretically: motor intervals $N$4 are independent of semantic content, whereas CLC couples $N$5 with $N$6, so forging the signal requires reproducing $N$7 rather than merely $N$8.

The framework further argues that attack cost is high. Even a single-session forgery is estimated to require 30–60 minutes of rehearsal per 1,500 words, combining memorization and practice, and multi-session consistency checks can further expose staged performances. This suggests that the rules are not merely classification heuristics but are intended as a threat-model-aware behavioral security protocol.

6. Implementation in Writing Interfaces

The implementation guidelines are designed for integration into existing editors or learning-management systems with minimal friction (Condrey, 26 Feb 2026). Instrumentation is to be performed by a lightweight plugin or browser extension that listens only to key-down timestamps and explicitly does not record key identity or screen content. Quantization is applied immediately on the client through $N$9, and raw high-resolution timings are discarded.

Phase labeling is maintained by a small in-memory state machine using the planning, translating, and revising thresholds already specified. For complexity estimation, the framework proposes computing a local score $t_i$0 when a token is committed, for example on space or punctuation, using an n-gram surprisal API or an on-device lightweight syntax depth estimator. CLC is then recomputed over a rolling window after each sentence or after a fixed number of keystrokes such as 200, and surfaced in a privacy-safe dashboard.

The privacy and control model is unusually explicit. The interface should provide an opt-in toggle, a persistent badge reading “Timing Collection On,” and a “View My Evidence” panel that displays only the aggregate CLC value and phase histogram, never raw timestamps. Once verification is completed, all quantized IKIs are deleted, retaining only the