Long-Range Dependence in Word-Length Series: A Graded Response Across Natural Text, Markov Surrogates, and Null Controls

An independent replication of long-range dependence in written natural language, extended by a control design that separates local structure from what lies beyond it. Long-range dependence in text — statistical influence persisting across distances far greater than any sentence — has been documented since the early 1990s and is well established. This study replicates it on word-length series using rescaled-range analysis, across 57 texts spanning six domains: novels, epic and myth, poetry, modern nonfiction, religion and philosophy, and science and technology. Languages include English, German, French, Latin, and Middle English. The extension is the control design. Rather than a single shuffle, the study uses two independently constructed control families. Shuffled controls destroy all order. Null sequences preserve the word-length distribution and nothing else. Markov surrogates preserve word-to-word transitions but no structure above the word pair. The response is graded rather than binary. Natural text gives a mean Hurst exponent of 0.6025. Markov surrogates give 0.5789. Null sequences give 0.5398 and shuffled natural text gives 0.5412 — two baselines built by different methods, not tuned to each other, agreeing to within 0.0014. All 57 texts exceeded their own shuffled controls without exception (paired t(56) = 12.42, p < 10⁻¹⁶, Cohen's d = 1.65). Local word-to-word transition structure accounts for approximately 64% of the effect. The remaining share does not follow from it. This decomposition is available only because two control families were used rather than one. The exponent proved invariant under concatenation, including concatenation of texts drawn from three different genres — evidence against a length artefact. Three polytonic Greek texts were excluded after their measurements saturated the estimator's ceiling; the analyser's tokeniser matched Latin script only. All three are named, their values reported in full, and the reason given. The complete 80-file manifest with every measurement appears in Appendix A and in the accompanying JSON. No claim is made about mechanism, meaning, coherence, or cognition. Hierarchical organisation is identified as a candidate explanation for the residual and marked explicitly as an untested hypothesis. This version supersedes the October 2025 white paper, which contained errors of priority, method description, and effect size. Those errors are enumerated in the paper and in the JSON rather than silently repaired. The experimental design required no correction.