FractalTextGuard: Structural Text Analysis via Long-Range Dependence

FractalTextGuard is an offline, dependency-free Python toolkit for structural text analysis using Long-Range Dependence (LRD) and Detrended Fluctuation Analysis (DFA). It is designed for reproducible auditing of long-form text by measuring correlation structure and compressibility patterns rather than stylistic keywords. Core metrics: - Hurst exponent (H) via DFA-based proxy (LRD signal) - Repetition rates (n-gram / line-level repetition signals) - Compression ratio (algorithmic redundancy proxy) - Shannon entropy (symbolic variability proxy) - Optional auxiliary stability signals (if enabled) Verdicts (policy-driven): - HEALTHY: no strong structural anomalies under configured thresholds - WARNING: ambiguous / borderline / low-confidence or mixed-structure cases - CRITICAL: strong structural anomalies consistent with high persistence and/or high redundancy Important limitations & minimum requirements: - Not reliable on short inputs. Recommended minimum length: >= 5,000–10,000 characters (preferably >= 20,000 for stable results). - Performance depends on text size and enabled options. For very large files (multi-MB), prefer chunked/batch runs and timeouts. - This tool provides structural indicators, not ground-truth authorship. It must not be used as the sole evidence to accuse individuals. ### Important limitations This is an experimental tool and **must not** be used as sole evidence in high-stakes decisions. Results are unreliable on short texts; human writing variation (including non-native writing or very structured styles) can increase false positives. Always combine with other checks and human review. Security / offline: - Fully offline: no network calls and no data leaves the host system. - Stdlib-only design is intended for “sealed” environments and reproducible forensic pipelines. This Zenodo release provides HEALTHY/WARNING/CRITICAL verdict mapping and an “industrial” workflow orientation (batch usage, deterministic seeds, and audit-friendly reporting).