Methodology
Which relationships between what is written in the Quran (words, concepts, letters, positions) and numbers survive an honest statistical screen, and which do not?
Pipeline
- Verified corpus: Tanzil 1.1 (Uthmani, Simple, Simple-Clean), hash-verified before parsing; 114 surahs, 6,236 verses; every line matched byte-for-byte against the XML export of the same edition.
- Named views: (edition, basmala policy, normalization profile) → token index with canonical offsets.
- The full search space is enumerated before anything runs (the denominator); every executed test is disclosed, including negatives.
- Screening: Benjamini–Hochberg over every executed test; a test becomes a “candidate pattern” only if q ≤ 0.05 and support ≥ 5.
- Deterministic: the same seed and inputs give the same result bytes; the run id is a hash of all inputs.
Run run-9a7c1a1988a98abf · seed 20260925 · engine 0.1.0 · commit d287965b5422
source tree sha256 8670eb30e4d3177be263f8600e78b1fe3030f9034a670ecfb0edf1c032d41211
results sha256 7810c7cf49bf9242806c01a4575854e23f20e79df86f23078807bca00aa82a9d
Families and null models
| Family | Question | Null model | Executed / candidates |
|---|---|---|---|
declared-number | Does a concept's count equal a declared number (e.g. day = 365, month = 12)? | Share of vocabulary forms with count in [N/2, 2N] whose count is exactly N | 6 / 0 |
concept-balance | Are two paired concepts' counts closer than chance? | Exact enumeration over all vocabulary pairs in the same band | 20 / 0 |
concept-cooccurrence | Do paired concepts share verses more than chance? | Hypergeometric over verses | 10 / 6 |
muqattaat-enrichment | Is an initial letter's share in its surah higher than in random verse sets? | Seeded verse resampling (9,999 draws), calibrated against the negative control | 1,174 / 0 |
muqattaat-specificity | Are initials enriched more than other letters of the same surahs? | One-sided Mann–Whitney U | 1 / 0 |
surah-concentration | Is a word form concentrated in one surah? | Binomial by surah word share, Bonferroni ×114 | 400 / 116 |
occurrence-clustering | Are occurrences burstier than random? | Seeded Monte-Carlo (999 uniform placements) | 101 / 71 |
Normalization profiles
Every text transformation is named, versioned and hashed; the published text is never altered.
| Profile | Description | Transforms | SHA-256 |
|---|---|---|---|
exact-v1 v1.0.0 | Exact surface Byte-exact token text including every diacritic and sign. | none | 543be835c22b44de… |
letters-only-v1 v1.0.0 | Letters only (orthography preserved) Drops every diacritic/sign; keeps the written letters exactly, incl. hamza forms and alef wasla. Default for letter counts. |
| 3b342c38f3efd772… |
search-folded-v1 v1.0.0 | Search-folded Letters only, dagger alef written out, alef wasla and hamza-alef forms folded to ا. Keeps ى/ي and ة/ه distinct. |
| 413547a18290fe13… |
search-folded-loose-v1 v1.0.0 | Search-folded (loose) search-folded-v1 plus ى→ي and ة→ه. Maximises recall; merges some distinct words. |
| 75d1ad678477723d… |
uthmani-no-harakat-v1 v1.0.0 | Uthmani without ordinary harakat Removes only U+064B–U+0652 (tanween, fatha, damma, kasra, shadda, sukun). Keeps Uthmani spelling, alef wasla, superscript alef, maddah, hamza marks, small waw/yeh, small Quranic signs and tatweel. |
| 1cb1685208dd821f… |
Basmala
Edition level: Tanzil keeps the surah-header basmala separate from verse text (XML export: a bismillah attribute on verse 1 of 112 surahs, all except 1 and 9). File-format level: the txt-2 export writes that attribute in-line at the start of the verse-1 line followed by one space. The metadata file (quran-data.xml) has no basmala field. The importer does not add or remove anything in the raw bytes; its policy is a named view.
Default policy: headers-removed — the surah-header basmala is not verse text (112 surahs); 1:1 and 27:30 are kept.
Concept counting rules
R1-bare— listed forms only, as whole tokens.R2-clitics— forms plus one uniform prefix set (و ف ب ل ك …), no suffixes.R3-substring— any token containing the stem (upper bound, includes unrelated words).
The same rules apply to every concept; a rule is never chosen per concept after seeing counts.
Claim ladder and AI policy
- The software stops at level 2 (statistically screened). Levels 3–6 (robustness across profiles, linguistic review, Quranic/statistical review, editorial publication) require human reviewers.
- AI proposed the concept forms and the two popular claims as hypothesis inputs labelled “draft, unreviewed”. It generated no Quran text, count, root or conclusion.
- No output uses “miracle” or “proof” wording; this is tested automatically.