{"preregistration":{"study":"RSA — Representational Similarity Analysis of the «اللسان المبين» hypothesis","registered_at":"2026-07-11","registered_by":"lisan research program (pre-registered by commit to git BEFORE any analysis ran; the registering commit hash is displayed on the dashboard)","background":"Lisan's core premise (L0) holds that the acoustic-articulatory energy signature of an Arabic root is not arbitrary with respect to its meaning. If true at corpus scale, the geometry of energy similarity across roots should share structure with the geometry of semantic similarity across the same roots. This is the doctorate-level protocol promised in the original Abu Awwad assessment: a falsifiable, corpus-wide test — not per-pair anecdotes.","hypotheses":{"H1":"Across the gold Quranic roots (n≈1,651), pairwise L0 energy similarity correlates positively with pairwise semantic similarity (Spearman Mantel ρ > 0).","H2":"H1 survives controlling for surface letter overlap: the partial Mantel correlation (controlling Jaccard similarity of root letter sets) remains positive and significant. This control is required because the semantic embedding text contains the root string and its derived surface forms, so shared letters could leak orthographic similarity into the semantic matrix."},"data":{"roots":"Qdrant collection lisan_roots — one point per gold root from the Quranic Arabic Corpus morphology (Dukes), rebuilt by research/reingest_roots.py","energy":"payload.energy_vector — 5 components (explosion, friction, resonance, flow, breath), computed by core.phonetics.root_energy_signature (letter-level energy vectors, emphatic amplification, averaged over the root's letters)","semantics":"the stored e5-large embedding (1024-dim) of «الجذر: X. السياقات: <up to 3 Quranic corpus contexts>»","letters":"payload.letters — the root's letter multiset, deduplicated to a set for the Jaccard control"},"exclusions":"Roots with missing/invalid energy_vector or fewer than 2 recognized letters are excluded; the excluded count is reported in the result.","method":{"similarity":"cosine similarity for both the energy matrix (5-dim) and the embedding matrix (1024-dim); Jaccard |∩|/|∪| for the letter-overlap control matrix","statistic":"Spearman rank correlation over the condensed upper triangle of the similarity matrices (n·(n−1)/2 pairs)","test":"Mantel permutation test: 10,000 random relabelings of the energy matrix (rows and columns permuted together); one-tailed p = (1 + #{ρ_perm ≥ ρ_obs}) / (10,000 + 1)","partial":"Smouse–Long–Sokal partial Mantel: partial Spearman ρ(AB·C) = (ρ_AB − ρ_AC·ρ_BC) / √((1−ρ_AC²)(1−ρ_BC²)), permutation applied to the energy matrix labels with ρ_BC held fixed"},"decision_rule":{"alpha":0.005,"min_effect":0.05,"supported":"p ≤ 0.005 for BOTH simple and partial Mantel AND partial ρ ≥ 0.05","detectable_but_negligible":"both p ≤ 0.005 but partial ρ < 0.05 — statistically real, practically weak; reported as such, not spun","confounded":"simple Mantel significant but partial Mantel not — the apparent effect is letter-overlap leakage; H1 is NOT supported","not_supported":"simple Mantel p > 0.005 (or ρ ≤ 0) — the corpus-wide version of the «اللسان المبين» premise fails this test"},"commitments":"The result is published on the public dashboard whichever way it falls. No post-hoc changes to similarity measures, exclusions, alpha, or effect threshold. Any exploratory follow-up analysis will be labeled exploratory."},"result":{"ran_at":"2026-07-11 17:14","runtime_s":438.1,"preregistration_sha256_16":"84a7d1ddf9a0a088","n_roots":1651,"n_excluded":0,"n_pairs":1362075,"n_permutations":10000,"rho_simple":0.0554,"p_simple":0.0001,"rho_partial":0.0312,"p_partial":0.0008,"rho_energy_letters":0.2909,"rho_semantic_letters":0.0881,"verdict":"detectable_but_negligible","exploratory_per_component_rho_vs_semantic":{"explosion":0.0325,"friction":0.0266,"resonance":0.0254,"flow":0.023,"breath":0.0106}}}