EXP-0006 — Consecutive-ayah lexical cohesion vs shuffled-order surrogates

completed Deterministic computation

Question

Do consecutive ayahs within a surah share vocabulary roots more than the same ayahs in shuffled order would?

Hypothesis

One-sided: mean root-set Jaccard over consecutive within-surah ayah pairs exceeds the within-surah order-shuffle null. Declared PRIMARY null: ayah-order-shuffle-within-surah-v1 (tests local ordering only). Declared SECONDARY null: ayah-order-shuffle-global-v1 (adds surah-membership signal). Metric, nulls, replicates, and base seed fixed before execution (D-044).

Counting policy

Root sets per ayah from morphology dataset quran-morphology-8f38b39 (words without root tags contribute nothing); consecutive pairs are within-surah only; pairs with empty root-set union are skipped and counted. Interpretation limited to ordered lexical cohesion under the declared shuffles; no authorship inference.

Results — permutation test vs declared surrogate class

Metric: mean root-set Jaccard over consecutive within-surah ayah pairs (1 empty-union pairs skipped).

Primary endpoint (declared in spec) (null: ayah-order-shuffle-within-surah-v1)

observed 0.06034 · null mean 0.03917 ± 0.000782 · null range [0.03709, 0.04176] over 1,000 surrogates · one-sided (greater) · p = 0.000999 (add-one estimator; 1/1001 is the floor) · base seed 20260732 · full null sample persisted in the result file

Secondary endpoint (null: ayah-order-shuffle-global-v1)

observed 0.06034 · null mean 0.02365 ± 0.000528 · null range [0.02197, 0.02545] over 1,000 surrogates · one-sided (greater) · p = 0.000999 (add-one estimator; 1/1001 is the floor) · base seed 20260733 · full null sample persisted in the result file

Inference scope

Consecutive ayahs share vocabulary roots beyond what their surah's own ayahs in shuffled order would produce (primary), and beyond globally shuffled placement (secondary, which additionally reflects that surahs are topically distinct). This describes ordered, locally cohesive composition. It supports no inference about authorship, and comparable cohesion should be expected in any coherent text; matched-text baselines are future work (§13.5).

Reproducibility record

Corpus: quran:hafs-kufan:v1
Method version: adjacency-cohesion-vs-null-v1
Results sha256: 4b89983d73cae81587a106b5c70c89fa…
Environment: {"platform":"Linux-6.18.5-x86_64-with-glibc2.39","python":"3.11.15"}
Ran: 2026-07-31T03:03:08Z → 2026-07-31T03:03:30Z

Re-run with python3 services/research-worker/scripts/run_experiment.py data/research/specs/EXP-0006.json. Input checksums are recorded in the result file.