Concept · active

Concept — Textual Kinematics

Textual KinematicsTK

# Textual Kinematics

## Definition

**Textual Kinematics (TK)** is the study of text as a dynamical system — the kinematic analysis of how text unfolds over time, treating language production as motion with measurable rhythm, velocity, and dynamics. It detects AI-generated versus human-generated text via the *temporal physics* of writing, on the principle that the underlying generative process leaves measurable signatures in the text's flow.

The central hypothesis (the **Classical-Jazz Hypothesis**): LLMs produce smooth, classical dynamics; humans produce erratic, jazz-like dynamics. The strongest signal — `adversarialDeviation` — measures inter-signal agreement (Δ=32.1 in the canonical benchmark).

## Coined by

Alexandru Mares (allemaar)

## First published

2025 (research method established); production detector (SYL) v7.

## Canonical artifact

- Project hub: [[GUIDE|TK project hub]] (`alm-os/Projects/TK - textualkinematics.org/`)
- Domain: [textualkinematics.org](https://textualkinematics.org)
- Production implementation: [[syl|SYL]] — stringyourlines.com
- Method papers: in development (planned formalization of the kinematic-analysis framework).

## Related concepts

- [[syl|SYL]] — production detector implementation
- [[younndai|YounndAI]] — sibling work
- [[elastic-automators|Elastic Automators]] — adjacent diagnostic vocabulary for the same generators TK analyzes
- Classical-Jazz Hypothesis
- adversarialDeviation
- AI text detection
- Information-theoretic analysis of generated text
- [[textual-kinematics|Cluster: Textual Kinematics]] — parent Cluster

## Why it matters

Most AI text detectors are classifier-based — they train on examples and become brittle as new generators ship. Textual Kinematics works differently: instead of pattern-matching surface features, it measures *generation-process invariants* (the temporal physics that any LLM-driven generator inherits from its sampling process). This makes TK harder to adversarially defeat by simply training on the detector's outputs, because the signal lives in the dynamics rather than the lexicon.

TK also opens a research surface that classifier-based detection cannot: it produces interpretable signals (rhythm, dynamics, Δ-deviation patterns) that can be analyzed, reasoned about, and connected to specific generative-architecture choices. The detection is a side effect of the analysis; the analysis is the contribution.

## Status

`active` — research and production both in continuous development.