# COUNCIL OF 120 — BENCH 3 · THE HADĪTH BANK
*Machine-scholar bench, 30 seats (Sunnī collections ×10 · Shīʿī collections ×10 · comparative ×10).
Labeled as such; human review invited, never waited for. **REVIEW AND COUNSEL ONLY — nothing edited.**
Sitting: 2026-09-04, 01:47–02:12 AST. Layers E1–E7 read at quiescence (mtime 01:26, read 01:47 — 21 min ≫ 3s, skill law 10).*

---

## VERDICT

**THE OCEAN IS SOUND AND THE ENGINES ARE HONEST — THE HEADLINE PROSE IS NOT.**
The bench independently re-derived the ocean from the sealed corpora and matched E1 **exactly on all 39 books and on the 68,921 total — zero delta**. E5's tier ladder is a *perfect partition* of the ocean (sums to 68,921, verified). E7's 9,962 pairs are **100% resolvable** and the CROWN (387 = Bukhārī 234 / Muslim 153 on 124 verses) and GOLDEN (1,330 on 418) re-derive exactly. E2's "full = ≥90%" law holds on **all 2,392 full pairs, zero violations**. Two of the bench's own alarms were **false and are withdrawn** (skill law 18 fired twice — the checker was wrong, not the engine).

But **five real defects survive into the synthesis**, three of them in `THE-ANSWERS.md` — the document a reader will actually see: a formula-substring class inflating the #2 headline verse, an unclassified mention-vs-exposition problem inflating the #1, a grade mis-attribution that breaks the grade-verbatim law at the headline, a declared absence that is partly *present*, and 226 rows of a **rijāl** work counted as ḥadīth. None is fatal. All are cheap to fix. **The bank is fit to build P32 on, once F1–F5 are corrected.**

---

## WHAT THE BENCH RAN (evidence trail)

| # | Probe | Script | Result |
|---|---|---|---|
| 1 | Schema-first probe, one row per container (law 2) | `bench3_probe.py` | sunni = dicts-of-dicts ×7; shia = dicts-of-**lists** ×32 ✅ |
| 2 | Independent recount, **all 39 books**, meta ignored | `bench3_recount.py` | 68,921 — **delta 0** ✅ |
| 3 | Field/grade census, graders named | `bench3_recount.py` | 12 graders; ṣaḥīḥayn `grades: null` |
| 4 | Arabic-presence, placeholders, truncation, 10-row sample | `bench3_deep.py` | 12 sentinel strings; 408 empty sunni `ar` |
| 5 | Dedupe: exact, tail-matn, cross-book | `bench3_deep.py` | Bukhārī 5.9% exact-tail redundancy |
| 6 | E2/E5/E7 structural read | `bench3_verify.py` | tier ladder partitions to 68,921 ✅ |
| 7 | **Content-gate: opened the exemplar ḥadīth** (law 12) | `bench3_open2.py`, `bench3_final.py` | 13/13 verified after law-18 correction ✅ |
| 8 | Basmalah-substring test | `bench3_decisive.py` | **187/244 of 1:3 match only inside the basmalah** |
| 9 | Mention-vs-exposition test | `bench3_decisive.py` | 112:1 = 92% liturgical; 87:1 = 96% |
| 10 | Full-class sweep, **all 2,392 pairs**, LCS coverage | `bench3_lcs.py` | 2,182 exact + 210 at 90–99%; **0 below 90%** ✅ |
| 11 | Near-duplicate clustering (8-gram, Jaccard ≥0.5) | `bench3_final.py` | Bukhārī collapse 6.9% |
| 12 | Category check on two anomalous Shīʿī "books" | targeted | **Ghaḍāʾirī = rijāl, not ḥadīth** |

Scripts in the session scratchpad; every number below is reproducible from them against `00-CORPUS/hadith.json` (sha `fbc8744a…`) and `hadith_shia.json`.

---

## PART A — WHAT VERIFIED (the bench's all-clears, stated as plainly as the alarms)

> *A false alarm costs trust exactly as much as a false all-clear. These are the all-clears, and the bench signs them.*

**A1 · THE RECOUNT IS EXACT.** ✅ VERIFIED. All 39 books, iterated from the row containers with `meta` deliberately ignored:

```
SUNNI  36,390  (E1: 36,390)  delta 0      SHIA  32,531  (E1: 32,531)  delta 0
OCEAN  68,921  (E1: 68,921)  delta 0      books 7 + 32  (E1: 7 + 32)
```
Every one of the 39 per-book counts matched to the row. The brief asked for 4 books recounted; the bench recounted **all 39** — it was cheaper than choosing.

**A2 · THE KUTUB AL-SITTAH IS COMPLETE.** ✅ VERIFIED — the bench's brief asked whether Muslim / an-Nasāʾī / at-Tirmidhī are absent. **They are not.** The 7 Sunnī books are `bukhari, muslim, nasai, abudawud, ibnmajah, tirmidhi, malik` — the full six plus the Muwaṭṭaʾ. `THE-ANSWERS.md:6` is correct as written.

**A3 · E5's TIER LADDER IS A PERFECT PARTITION.** ✅ VERIFIED — 17,726 + 16,024 + 15,152 + 14,805 + 2,628 + 2,533 + 53 = **68,921 exactly**. Independently: T1 = 7,589 + 7,563 = 15,152 ✅; T4 "ungraded-canonical" = 53 = the bench's own recount of ungraded rows outside the ṣaḥīḥayn (ibn Mājah 2 + Nasāʾī 7 + Tirmidhī 44) ✅. No row is double-counted or orphaned.

**A4 · E7 IS FULLY RESOLVABLE AND ARITHMETICALLY CLEAN.** ✅ VERIFIED — 9,962 pairs, **0 unresolvable ḥadīth references**; 6,000 distinct carriers (E2 says 6,000 ✅); 2,350 distinct verses (E2 says 2,350 ✅). CROWN re-derived = **387**, by book **{bukhari: 234, muslim: 153}** on **124 verses** — matches `THE-ANSWERS.md:25` to the digit. GOLDEN re-derived = **1,330 on 418 verses** — matches line 26.

**A5 · THE "FULL QUOTE" TIER HONORS ITS OWN ≥90% LAW, EXACTLY.** ✅ VERIFIED — the bench swept **all 2,392** full-class pairs: 2,182 are exact substrings; the other 210 measure **90.0%–99% longest-common-substring coverage**; **not one falls below 90%**. This is the single strongest result of the sitting.

**A6 · TWO BENCH ALARMS WITHDRAWN (skill law 18).** The bench's first content-gate reported 4/9 exemplar failures. **The checker was wrong.** `quran.json` prepends the basmalah to every sūrah-opener's `uthmani` field; E2 correctly strips it, the bench's checker did not. Re-run like-for-like: **13/13 pass**. This is the 87:1 incident of law 18 recurring verbatim, and it vindicates the engine a second time. Separately the bench suspected Muḥsinī's *Muʿjam al-Aḥādīth al-Muʿtabara* (a modern digest) of reprinting al-Kāfī: measured **0 of 555** duplicated in any held book — **alarm withdrawn, no defect**.

**A7 · CROSS-BOOK CONTAMINATION IS NEGLIGIBLE.** ✅ VERIFIED — Sunnī: 4 cross-book exact duplicates in 35,566 distinct texts. Shīʿī: 4 in 31,036. The 39 books are genuinely 39 books.

---

## PART B — FINDINGS

### 🔴 F1 · CRITICAL — verse 1:3 is a substring of the basmalah; 187 of its 244 pairs are formula artefacts
**File:** `10-WAQF/build/LENS-AUDIT/E7_pair_registry.json` (class `full`, v=`1:3`) · surfaced at `THE-ANSWERS.md:19`
**Evidence** (`bench3_decisive.py`, section b):
```
basmalah skeleton : بسماللهالرحمنالرحيم
1:3      skeleton : الرحمنالرحيم     -> 1:3 INSIDE basmalah ? True
1:3  pairs=244  carriers containing basmalah=193
     carriers where the match EXISTS ONLY INSIDE a basmalah occurrence = 187  (77%)
```
Opened carriers: `abudawud#784` — *"A sūrah has just been revealed to me. He then recited: 'In the name of Allah, the Compassionate, the Merciful. Lo! We have given thee Abundance'"* — that is **Sūrat al-Kawthar's basmalah**, scored as a full quote of *al-Fātiḥah* 1:3. Likewise `#786`, `#787`, `#788` (all four are reports *about* the basmalah between sūrahs) and `#2879` (ʿUmar's **waqf deed**, which simply opens with the basmalah).

E2's law states *"the embedded basmalah is stripped from every opener except 1:1 so it forges nothing."* That shield runs one direction only — it strips the basmalah from the **ḥadīth**. It cannot help when the **verse itself is a substring of the formula**. 1:3 is 12 rasm characters long and sits wholly inside the basmalah's 19.
**Impact:** 1:3 is the **#2 headline verse** in `THE-ANSWERS.md:19` ("112:1 = 249 · 1:3 = 244"). Its honest evidentiary count is **≈57, not 244**. E7 is declared the P32 rebuild seed, so 187 bad pairs would propagate into the new pillar.
**Severity: CRITICAL** (headline number + seed contamination).
**Fix:** extend the formulaic law with its converse — **THE CONTAINED-FORMULA RULE**: *a verse whose entire skeleton is a substring of a high-DF liturgical formula may only be matched where the match survives deletion of every occurrence of that formula from the carrier.* One line of code; the bench has already computed the surviving count.

### 🔴 F2 · HIGH — the uncapped headline counts are liturgical mentions, not exegesis; the corpus has no MENTION/EXPOSITION class
**File:** `E7_pair_registry.json` · surfaced at `THE-ANSWERS.md:19`
**Evidence** (`bench3_decisive.py`, section c) — share of carriers whose English carries recitation/liturgical vocabulary:
```
112:1  pairs=249  liturgical 228 (92%)     87:1  pairs= 69  liturgical 66 (96%)
1:3    pairs=244  liturgical 161 (66%)     6:163 pairs= 20  liturgical 20 (100%)
```
All 8 carriers of 112:1 opened by hand are recitation reports, not exegesis: *"he offered the sunset prayer behind Ibn Masʿūd, when he recited 'Say: He is Allah, the One'"* (`abudawud#815`); *"the Prophet ﷺ would recite in the two rakʿahs of the dawn prayer…"* (`#1256`); *"every night when he went to his bed… reciting into them"* (`#5056`). **Zero of eight explain the verse's meaning.**
**Impact:** "112:1 has 249 links" is true as a *quote census* and misleading as an *exegetical* one. The waqf's own doctrine already says "a quote LOCATES; the reading bench EXPLAINS" (`THE-ANSWERS.md:57`) — but that caveat lives in the footer while the number lives in the headline. A reader will take 249 as 249 explanations of *al-Ikhlāṣ*.
**Severity: HIGH** (framing of the flagship number).
**Fix:** add a fourth axis beside class/tier: **`use` ∈ {recitation-report, ṣalāh-context, faḍāʾil, occasion-of-revelation, exposition, legal-inference}**. Mechanically seedable today from the English (92%/96% recall demonstrated above), then reviewed. Report the headline as *"112:1 — 249 quote-pairs, of which 21 non-liturgical."*

### 🟠 F3 · HIGH — the Shīʿī grades are attributed to one grader; there are **three**, across **twelve** books
**File:** `E1_ocean_census.json` → `honest_gaps[0]` · repeated at `THE-ANSWERS.md:28`
**Claim:** *"al-Majlisī's Mirʾāt al-ʿUqūl classifications ride al-Kāfī's 8 volumes verbatim"* / *"Majlisī-graded (Mirʾāt al-ʿUqūl, held verbatim on al-Kāfī) — 14,805 rows."*
**Measured** (`bench3_deep.py`, section 6):
```
graders: al-Majlisi 13,941 · Behbudi 1,966 · Mohseni 825
books carrying >=1 grade: 12 of 32  (not 8)
  al-Kafi vols 1-8 = 13,981 rows
  + Man-La-Yahduruh-Vol-5  390/395 (98.7% graded)
  + Mujam-al-Ahadith-al-Mutabara-Muhsini  428/555
  + Al-Khisal 4 · Uyun-akhbar-al-Rida-Vol-1 2
  = 14,805 rows total
```
So **824 graded rows sit outside al-Kāfī**, and **Behbudī (2,791 combined with Muḥsinī) is unnamed in the synthesis**. This is a direct breach of the grade-verbatim law (#9: *"grading strings are transmitted verbatim **with the grader named by the source**"*) at the most-read line in the audit.
**Credit where due:** **E5 has this right** — `graders_named_in_the_corpus` lists all twelve including Behbudi and Mohseni. The defect is the synthesis layer contradicting a correct layer beneath it.
**Severity: HIGH** (own-law breach, headline).
**Fix:** `THE-ANSWERS.md:28` → *"Shīʿī graded rows **14,805 across 12 books**, graders named: al-Majlisī 13,941 (Mirʾāt al-ʿUqūl, on al-Kāfī) · Behbudī 1,966 · Muḥsinī 825."*

### 🟠 F4 · HIGH — a declared absence that is partly present: rijāl data **is** held, and 226 rows of it are miscounted as ḥadīth
**File:** `E1_ocean_census.json` → `honest_gaps[1]` · repeated at `THE-ANSWERS.md:31`
**Claim:** *"per-narrator rijāl data (Taqrīb al-Tahdhīb class, asmāʾ al-rijāl) is **NOT held**."*
**Measured:** `Kitab-al-Duafa-Ghadairi` (226 rows) is **Ibn al-Ghaḍāʾirī's Kitāb al-Ḍuʿafāʾ — a rijāl work**, and its rows are narrator entries, not narrations:
```
"1 – Abān b. Abī ʿAyyāsh … A tābiʿī, he narrated from Anas b. Mālik … Weak, he is not to [be relied upon]"
"2 – Ibrāhīm b. ʿUmar, al-Ṣanʿānī, al-Yamānī … Very weak. He narrated from Abū Jaʿfar …"
```
Two defects at once: **(a)** the absence is over-declared — a foundational Imāmī jarḥ work *is* in hand, and the "20 knowledge areas" gate has a seed today; **(b)** 226 narrator-evaluations are counted inside the 68,921 "ḥadīth". Related: `Risalat-al-Huquq-Abidin` (49) is **one treatise** sectioned into 49 rows, and `Nahj-al-Balagha` (2,260) is khuṭab/letters/aphorisms of ʿAlī — legitimately in a ḥadīth bank under Imāmī usage, but not prophetic ḥadīth.
**Severity: HIGH** — I-3 cuts both ways; a false all-clear about an *absence* would send an acquisition team to buy what is already on the shelf.
**Fix:** add a `genre` field per book (`hadith | rijal | khutab | risala`); report the ocean as **68,921 rows / 68,646 narrations / 226 rijāl entries / 49 treatise sections**.

### 🟠 F5 · HIGH — the ṣaḥīḥayn carry **no grade field**; CROWN is a collection-inference, presented under a grade-verbatim heading
**File:** `hadith.json` · surfaced at `THE-ANSWERS.md:22–25`
**Evidence:** `bukhari: NoneType:empty ×7,589` and `muslim: NoneType:empty ×7,563` — **15,152 rows (41.6% of the Sunnī ocean, 22.0% of the whole) carry `grades: null`.**
`THE-ANSWERS.md:22` opens the tier table with *"'Exact meaning' has a ladder … (grade-verbatim law)"* and its top rung is **CROWN = full-verse quote × Bukhārī/Muslim = 387**. But for those two books there is **no transmitted grade string to be verbatim about**. CROWN rests on *collection identity* (the ummah's ittifāq on the ṣaḥīḥayn) — a defensible and traditional inference, but a **different kind of claim** than "al-Albānī said ṣaḥīḥ", and the reader is not told which they are holding.
**Credit where due:** **E5 names it correctly** — the rung is literally `"T1 collection-sahih"`. Again the synthesis is weaker than the layer.
**Severity: HIGH** (the flagship tier's epistemic basis is unlabelled).
**Fix:** rename the CROWN row *"full-verse quote × ṣaḥīḥayn (**standing by collection — the corpus transmits no grade string for Bukhārī/Muslim**)"*. Cost: one parenthesis. Gain: the strongest claim in the audit becomes unattackable.

### 🟡 F6 · MEDIUM — coverage honesty: E1 names **zero** absent collections, and half the Shīʿī Four Books are missing
**File:** `E1_ocean_census.json` → `honest_gaps` (3 entries, none about missing books)
**Measured — what is held:** al-Kāfī (8 vols, 14,245) + Man Lā Yaḥḍuruh (5 vols, 6,382) = **2 of the 4** *al-Kutub al-Arbaʿa*.
**Absent (Shīʿī):** **Tahdhīb al-Aḥkām** and **al-Istibṣār** — Ṭūsī's two, i.e. **half the Shīʿī canon**; also Wasāʾil al-Shīʿa, Biḥār al-Anwār, Mustadrak al-Wasāʾil, Baṣāʾir al-Darajāt, Amālī al-Ṭūsī.
**Absent (Sunnī, beyond the sittah):** Musnad Aḥmad, Sunan al-Dārimī, Ṣaḥīḥ Ibn Ḥibbān, Ṣaḥīḥ Ibn Khuzayma, al-Mustadrak, al-Sunan al-Kubrā (Bayhaqī), the two Muṣannafs, Ṭabarānī's Muʿjams.
`THE-ANSWERS.md:54` *does* name most of these — so the commission knows. The defect is that **E1, the layer a machine will read, declares three gaps and not one of them is "which books are missing"**. A layer that does not state its own boundary invites the next engine to treat 68,921 as the ḥadīth corpus.
**Severity: MEDIUM.** **Fix:** add `honest_gaps[3]`: the absent-collections list, with the Four-Books gap named first.

### 🟡 F7 · MEDIUM — the Shīʿī `matn` and `sanad` fields contain **English**, not Arabic
**File:** `hadith_shia.json` — every book
**Evidence** (`bench3_decisive.py`, section e): of 32,531 rows, `matn` is **Arabic-dominant in 2** and **Latin/English-dominant in 27,228** (5,301 empty). `ar == sanad + matn` in **0 of 32,531** rows. Sample:
```
ar   : ١ - حدثنا الشيخ الأجل المفيد أبو عبد الله محمد بن محمد بن النعمان …
matn : "1. The grand Shaikh al-Mufid Abu Abdillah Muhammad ibn Muhammad ibn al-Nu'man …"
```
**Impact:** a field named `matn` that is not the Arabic matn is a **silent-zero landmine** — a P32 engine that reasonably matches Arabic against `matn` returns 0 hits and looks merely "sparse". This is precisely the failure class law 2 exists to prevent, one layer deeper. E2 correctly used `ar`; the next engine may not.
**Severity: MEDIUM** (zero impact today, high probability of a silent future defect).
**Fix:** document in the P32 schema note: *"`matn`/`sanad` are the **English** translation's split. There is no Arabic isnād/matn separation in either corpus."*

### 🟡 F8 · MEDIUM — no Arabic isnād/matn split exists in **either** corpus; all quote matching is isnād-blind
**Evidence:** Sunnī rows have no such fields at all (`ar, book, hadith_in_book, en, ur, grades`); Shīʿī rows have them in English only (F7). Therefore all 68,921 match surfaces are **isnād + matn fused**, and every narrator name — *ḥaddathanā … ʿan … qāla* — sits inside the searchable text.
**Consequence for the counsel:** the tradition's own unit of analysis is the **matn**, and the waqf cannot currently address it. This is the root cause of F9 and the reason a narration-family count is impossible today.
**Severity: MEDIUM.** **Fix:** an isnād-head stripper (the `ḥaddathanā/akhbaranā/ʿan X ʿan Y qāla` prefix is highly regular) as a named P32 engine, validated against the 27,228 English matn boundaries the Shīʿī corpus already provides as ground truth. **That English split is free supervised training data — the bench recommends exploiting it.**

### 🟡 F9 · MEDIUM — dedupe reality: links are counted per **row**, not per **narration**; Bukhārī's mukarrar is invisible
**Evidence** (`bench3_deep.py` §3, `bench3_final.py`):
```
exact full-text duplicates, in-book : bukhari 297 groups / 608 rows (redundancy 311)
exact matn-TAIL (last 150 chars)    : bukhari 5.3%  muslim 1.3%  malik 0.0%
near-duplicate, 8-gram Jaccard>=0.5 : bukhari 6.9%  muslim 1.3%  malik 0.2%
largest family: bukhari 958/959/960/961 — same text ×4
```
Bukhārī is a 5–7× outlier — the mukarrar signature is real and detectable. But the tradition's own figure (≈7,400 with repetition vs ≈2,600 without ⚠️ ESTIMATE, standard editions) implies **~65%** redundancy at the narration level. The gap between the measured 6.9% and that ~65% is **variant** repetition — same matn, different isnād and different wording — which exact and Jaccard-0.5 methods cannot see, and which F8 makes structurally unreachable.
**Consequence:** "112:1 has 249 quote-pairs" counts **rows**. The number of *distinct narrations* behind them is unknown and certainly smaller.
**Severity: MEDIUM** (no number is wrong; a number is missing).
**Counsel — the bench's proposed policy, for the record:**
> **THE NARRATION-FAMILY POLICY.** Rows are never merged or deleted (I-6). Every pair keeps its row identity. A **family id** is *added* — rows whose matn-skeletons cluster at Jaccard ≥ 0.8 after isnād-stripping share a family. Every link count is then published as a **pair**: *"249 pairs / N narration-families"*. Ambiguity is data: a row that clusters into two families is recorded in both, never forced. **Never** collapse across collections — Bukhārī's and Muslim's transmissions of one matn are two witnesses, and merging them would destroy the very corroboration (`mutābaʿa`) the ladder is built to measure.

### 🟢 F10 · LOW-MEDIUM — 12 Shīʿī rows carry **sentinel placeholders** in the Arabic field
**Evidence** (`bench3_deep.py` §1): `'To be added'` ×2, `'ArabicToAdd'`, `'arabicnotfound'`, `'notfound'` ×3, `'-'` ×3, `'s'`, `'* * * * *'` — in Kāmil al-Ziyārāt, Kitāb al-Ghayba (Ṭūsī), Kitāb al-Zuhd, Nahj al-Balāgha #2071, Thawāb al-Aʿmāl.
**Impact:** these strings are **non-empty** and pass any `if row['ar']:` gate — they are worse than nulls, because they are invisible to the obvious check. Zero impact on E2 (they cannot match a verse), but they will silently pollute any future completeness metric.
**Severity: LOW-MEDIUM.** **Fix:** a sentinel blocklist in the ingest gate; count them as missing, never as text.

### 🟢 F11 · LOW-MEDIUM — 408 Sunnī rows have empty Arabic, **203 of them in Ṣaḥīḥ Muslim**, and no layer says so per book
**Evidence:** `muslim 203 · nasai 86 · tirmidhi 74 · malik 29 · bukhari 9 · ibnmajah 5 · abudawud 2 = 408`. Sample row `muslim#1` is **wholly empty** — `ar_len=0, en_len=0, grades=null`.
**Impact:** 2.7% of Ṣaḥīḥ Muslim can never be quote-matched — and Muslim is half of the CROWN tier (153 of 387). The CROWN is therefore computed over 7,360 of 7,563 Muslim rows. E1 reports only the aggregate `rows_with_arabic: 35,982`; the per-book holes are undeclared.
**Severity: LOW-MEDIUM.** **Fix:** publish `rows_with_arabic` per book; note the Muslim deficit wherever CROWN is stated.

### 🟢 F12 · LOW — E2 scanned 68,505 of 68,921; the 416-row gap is not declared
**Evidence:** `E2.totals.hadith_scanned = 68505` vs the measured ocean 68,921 → **416 rows never entered the quote engine** (the empty/sentinel rows of F10–F11, near enough). E2's `totals` block does not state this.
**Severity: LOW** (absences must be declared, law 9 / I-1). **Fix:** one field, `rows_not_scanned: 416, reason: "no Arabic text"`.

### 🟢 F13 · LOW — three definitions of "has Arabic" are in play; E1's is undocumented
**Evidence:** non-empty string → **32,531**; contains an Arabic letter → **32,519**; E1 reports → **32,523**. All three are defensible; none is written down. The bench does **not** call E1 wrong — it calls the definition unstated.
**Severity: LOW.** **Fix:** state the predicate beside the number.

### 🟢 F14 · LOW — 43 rows end in an ellipsis (real truncation), all Shīʿī
**Evidence:** Al-Khiṣāl #468, Ghayba (Ṭūsī) #127, Ghayba (Nuʿmānī) #243, Muʿjam (Muḥsinī) #69 and #96, + 38 more. **Severity: LOW** (0.06%), but they are *silently* short and will under-match. Flag, do not repair.

### 🟢 F15 · INFORMATIONAL — the "full quote" tier leans on very short verses
**Evidence** (`bench3_final.py` extension): of 2,392 full pairs, **707 (29.6%) rest on verses of ≤12 rasm characters**, and **two verses — 112:1 and 1:3 — supply 493 of those 707 (70%)**. Sixty-seven distinct verses are that short.
**Why it matters:** F1 is the *instance*; this is the *class*. Short verses are structurally prone to being contained inside longer formulae, names and refrains. **The bench recommends a standing length-aware rule**: below ~15 rasm characters, a full-quote claim requires a second signal (context window, sūrah-name adjacency, or a non-formulaic neighbour gram).

### 🟢 F16 · INFORMATIONAL — the grade layer is overwhelmingly **modern**, and this is nowhere said
**Evidence:** Al-Albānī 19,081 · Zubair Ali Zai 19,033 · Shuʿayb al-Arnaʾūṭ 6,413 — the three dominant graders are **20th-century**. Classical-edition annotators (ʿAbd al-Bāqī 4,314, Aḥmad Shākir 3,729, Bashshār ʿAwwād 2,425, Abū Ghudda 5,692, Muḥyī al-Dīn ʿAbd al-Ḥamīd 5,171, Sālim al-Hilālī 1,858) are also modern editors. **No pre-modern Sunnī grader appears in the corpus at all.**
The grade-verbatim law is *satisfied* — every grader is named. But a reader who sees "ṣaḥīḥ" on a tier ladder will assume a classical verdict. **Severity: INFORMATIONAL, high communicative value.** **Fix:** one line under the ladder — *"grading strings are modern (Albānī, Zubair ʿAlī Zaʾī, al-Arnaʾūṭ and modern editors); the corpus transmits no pre-modern Sunnī grade field."* This *strengthens* the waqf's honesty claim rather than weakening it.

---

## MECHANICAL TABLE

| # | Check | Bench measurement | E1/E2/E5/E7 claim | Δ | Label | Severity |
|---|---|---|---|---|---|---|
| 1 | Ocean total | **68,921** | 68,921 | **0** | ✅ VERIFIED | — |
| 2 | Sunnī rows / books | **36,390 / 7** | 36,390 / 7 | 0 | ✅ VERIFIED | — |
| 3 | Shīʿī rows / books | **32,531 / 32** | 32,531 / 32 | 0 | ✅ VERIFIED | — |
| 4 | Per-book counts, all 39 | **all match** | — | 0 | ✅ VERIFIED | — |
| 5 | Sunnī rows w/ Arabic | **35,982** | 35,982 | 0 | ✅ VERIFIED | — |
| 6 | Shīʿī rows w/ Arabic | 32,531 / **32,519** | 32,523 | ±8 | 📊 CALCULATED | F13 LOW |
| 7 | Sunnī grade rows | **21,185** | 21,185 | 0 | ✅ VERIFIED | — |
| 8 | Shīʿī grade rows | **14,805** | 14,805 | 0 | ✅ VERIFIED | — |
| 9 | Shīʿī graded **books** | **12 of 32** | "the other **24** carry no grade" | **−4 books / 824 rows** | ✅ VERIFIED | F3 HIGH |
| 10 | Shīʿī graders | **3** (Majlisī/Behbudī/Muḥsinī) | "Majlisī" (E1/ANSWERS); **12 named (E5 ✅)** | 2 unnamed in synthesis | ✅ VERIFIED | F3 HIGH |
| 11 | ṣaḥīḥayn grade fields | **0 of 15,152** | "T1 collection-sahih" (E5 ✅) | synthesis unlabelled | ✅ VERIFIED | F5 HIGH |
| 12 | Tier ladder partition | **sums to 68,921** | 68,921 | 0 | ✅ VERIFIED | — |
| 13 | T4 ungraded-canonical | **53** (2+7+44) | 53 | 0 | ✅ VERIFIED | — |
| 14 | E7 resolvability | **0 unresolvable / 9,962** | — | 0 | ✅ VERIFIED | — |
| 15 | CROWN | **387** (Bukh 234 / Musl 153), 124 verses | 387, 124 | 0 | ✅ VERIFIED | — |
| 16 | GOLDEN | **1,330 on 418** | 1,330 on 418 | 0 | ✅ VERIFIED | — |
| 17 | Full-class ≥90% law | **2,392/2,392 pass; 0 below 90%** | ≥90% | 0 | ✅ VERIFIED | — |
| 18 | Exemplar content-gate | **13/13** (after law-18 fix) | — | 0 | ✅ VERIFIED | A6 |
| 19 | **1:3 basmalah artefacts** | **187 of 244 (77%)** | 244 headline | **−187** | ✅ VERIFIED | **F1 CRITICAL** |
| 20 | **112:1 liturgical share** | **228 of 249 (92%)** | 249 headline | unclassified | ✅ VERIFIED | **F2 HIGH** |
| 21 | 87:1 liturgical share | **66 of 69 (96%)** | 69 | unclassified | ✅ VERIFIED | F2 HIGH |
| 22 | Ghaḍāʾirī = rijāl not ḥadīth | **226 rows** | "rijāl NOT held" | contradicted | ✅ VERIFIED | F4 HIGH |
| 23 | Shīʿī `matn` language | **27,228 English / 2 Arabic** | field named `matn` | — | ✅ VERIFIED | F7 MED |
| 24 | Bukhārī redundancy (exact-tail) | **5.9%** | not measured | — | 📊 CALCULATED | F9 MED |
| 25 | Bukhārī redundancy (near-dup J≥0.5) | **6.9%** | not measured | — | 📊 CALCULATED | F9 MED |
| 26 | Bukhārī mukarrar, tradition's own | **~65%** | not measured | unreachable (F8) | ⚠️ ESTIMATE | F9 MED |
| 27 | Muḥsinī reprints al-Kāfī? | **0 of 555** | — | — | ✅ VERIFIED | **alarm withdrawn** |
| 28 | Cross-book dups (S / Sh) | **4 / 4** | — | — | ✅ VERIFIED | — |
| 29 | Sentinel placeholder rows | **12** | undeclared | — | ✅ VERIFIED | F10 LOW-MED |
| 30 | Empty Arabic, Ṣaḥīḥ Muslim | **203 (2.7%)** | aggregate only | — | ✅ VERIFIED | F11 LOW-MED |
| 31 | E2 rows never scanned | **416** | undeclared | — | ✅ VERIFIED | F12 LOW |
| 32 | Ellipsis-truncated rows | **43** | undeclared | — | ✅ VERIFIED | F14 LOW |
| 33 | Full pairs on verses ≤12 chars | **707 (29.6%)** | — | — | 📊 CALCULATED | F15 INFO |
| 34 | Pre-modern Sunnī graders | **0** | — | — | ✅ VERIFIED | F16 INFO |
| 35 | Kutub al-sittah complete? | **YES — all 6 + Muwaṭṭaʾ** | "full sittah" | 0 | ✅ VERIFIED | — |
| 36 | Shīʿī Four Books complete? | **NO — 2 of 4** (Tahdhīb, Istibṣār absent) | unnamed in E1 | — | ✅ VERIFIED | F6 MED |

---

## GAP-HIGHLIGHT DUTY — what the commission itself missed
*(skill law 13; beyond the five findings above, which are themselves misses)*

1. **The converse of the formulaic law was never written.** E2 shields against *formulae inside ḥadīth*. Nobody asked about *verses inside formulae*. F1 is one instance; F15 shows 67 verses in the exposure class. The law is half-built.
2. **Presence was measured; USE was never measured.** Every layer answers *"is the verse in this ḥadīth?"* No layer asks *"why is it there?"* — and 92% of the time the answer is "he recited it in prayer." The four-witness doctrine implicitly promises exegetical witness and currently delivers citation witness. **This is the largest single gap in the audit.**
3. **The ocean was counted; it was never classified by GENRE.** 68,921 mixes prophetic narrations, imāmī akhbār, khuṭab, a legal treatise and a **rijāl dictionary** in one number (F4). Every downstream ratio inherits the blend.
4. **A held asset was declared absent.** Ibn al-Ghaḍāʾirī's *Ḍuʿafāʾ* is on the shelf while the plan budgets to acquire rijāl data. Nobody opened the smallest books. **The bench's rule: read the shortest book in the corpus first — it is the one nobody has ever opened.**
5. **The English translations are an unexploited ground-truth layer.** 27,228 Shīʿī rows carry a human-made isnād/matn boundary (F7). The waqf treats this as a schema quirk; it is free supervised data for the isnād-stripper that F8/F9 require.
6. **"Uncapped" was celebrated without asking what the cap was hiding.** Removing the cap of 9 was correct and is a real win. But the new top of the distribution (112:1, 1:3, 87:1) turns out to be liturgy and basmalah — i.e. **the cap was accidentally suppressing the artefact class**. Uncapping was right; it also removed a shield nobody knew was there.
7. **No layer states the corpora's provenance chain.** `thaqalayn-api.net` is in the Shīʿī meta; the Sunnī meta names no source at all. For a waqf whose whole claim is transmitted honesty, **the corpus's own isnād is undocumented.** MANIFEST.sha256 seals the bytes but not their origin.
8. **Nobody checked whether the ocean is growing in the right dimension.** Adding 51,000 rows of Musnad Aḥmad multiplies *rows*; it may barely move *verses covered*, which is the metric that matters (E6: 702 verses have neither ḥadīth nor sīrah). **The bench recommends every acquisition be scored on Δ-verses-covered, not Δ-rows** — see Q5.

---

# COUNSEL

*Thirty seats; votes cast by sub-bench. Dissents preserved verbatim, never flattened.*

## Q1 — the split of the 3,600 free nodes (26,400 → 30,000)

### 🗳 BENCH VOTE: **(b) — a modified split. 26 for · 4 dissenting.**

| Allocation | Option (a) | **Bench (b)** | Δ |
|---|---|---|---|
| P28 Sīrah Universe 300 → 1,700 | +1,500 | **+1,400** | −100 |
| **P32 (new pillar)** | +1,800 | **+1,800** | 0 |
| **Acquisitions & rijāl reserve** | +300 | **+400** | **+100** |
| **Total** | 3,600 | **3,600** | — |

**One-line reason:** the bench measured the acquisition workload it is voting for in Q5 — **four collections, ≈+50,251 rows, a 73% ocean expansion, two ingest pipelines, a grade-field mapping for two ungraded Ṭūsī books, and a full re-run of E2/E5/E7 over a 119k ocean** — and 300 nodes cannot carry it; the 100 nodes come from Sīrah, which still receives a 4.7× expansion and is not the bench's field to defend.

**Why P32 is *not* increased despite this bench producing the most findings:** F1, F2, F9 and F15 are each one law or one classifier — they are *cheap*, and the bench refuses to inflate a budget on the strength of its own findings. Honest cost, not advocacy.

> **DISSENT — Seats C7, C9, C10 (comparative) and S4 (Sunnī):** *hold the reserve at +300 and give the 100 to P32.* Reason: acquisition is procurement and can be re-budgeted when it is actually scoped; F2's `use` classifier will need human review at 9,962-pair scale and that is where the nodes will really go. **Recorded, not adopted.**

## Q2 — P32 as a NEW pillar vs an expansion of the Hadīth Universe; and its NAME

### 🗳 BENCH VOTE: **NEW PILLAR — 30 for · 0 against.** *(the only unanimous vote of the sitting)*

**One-line reason:** P32's subject is **the relation between witnesses**, not ḥadīth — it carries the cross-verse witness (`tafsīr al-Qurʾān bi-l-Qurʾān`, 6,069 pairs) and the sīrah joins, and filing the **Qurʾān's witness to itself** under a *Ḥadīth* heading is a category error the tradition would not forgive.

**Two supporting reasons the bench records:** (i) **its dependency runs the wrong way** — P32 *consumes* the Ḥadīth Universe, and a layer that consumes a pillar cannot live inside it without a cycle; (ii) **it must be able to contradict** — F1/F2/F9 are corrections *about* the ḥadīth bank, and a bench cannot audit its own parent.

### 🗳 NAME: **“THE FOUR WITNESSES” (P32)** — 22 for · 8 for the Arabic alternative.

**⚠️ The bench raises a technical objection nobody else is positioned to raise.** The obvious Arabic name — *Majmaʿ al-Shawāhid* — **collides with established ḥadīth terminology**: a *shāhid* is a corroborating narration from a **different Companion**, and *mutābiʿ* from the same one. Naming the four-witness pillar *al-Shawāhid* would make "shāhid" mean two different things inside one waqf, and the collision would surface precisely in P32's own corroboration engine. **The bench recommends the plain English name and rejects *al-Shawāhid*.**

> **MINORITY (8 seats, Shīʿī sub-bench H2–H9):** prefer ***Pillar of the Four Witnesses — Arkān al-Shahāda al-Arbaʿa***, arguing an Arabic name signals the waqf's register. **Recorded.** The bench notes *shahāda* carries its own load (testimony/martyrdom/the creed) and that the majority's caution applies with equal force.

## Q3 — the four-witness doctrine wording

### 🗳 BENCH VOTE: **APPROVE WITH FOUR AMENDMENTS — 28 for · 2 for approve-as-written.**

The doctrine (`SKILL.md:48–55`) is sound in structure. The bench's amendments are all **consequences of what it measured today** — none is stylistic.

**AMENDMENT 1 — to witness (2), the ḥadīth witness.** Current: *"the HADITH (graded, carriers named)."* **The parenthesis is false for 15,152 rows** (F5). Amend to:
> *(2) the ḤADĪTH — **carriers always named; graded where the corpus transmits a grade string, and where a report's standing rests on its collection rather than on a transmitted grade (the ṣaḥīḥayn), that is declared as collection-standing, never presented as a grade).***

**AMENDMENT 2 — a new closing clause, THE LOCATION CLAUSE.** The distinction exists in `THE-ANSWERS.md:57` but not in the doctrine, and F2 shows it is the doctrine's most load-bearing sentence:
> *A witness may be **present** without being **exegetical**. A quote LOCATES; it does not yet EXPLAIN. Every link therefore carries its **use** — recitation-report, ṣalāh-context, faḍāʾil, occasion-of-revelation, exposition, legal-inference — and a count of witnesses is never reported as a count of explanations.*

**AMENDMENT 3 — to witness (4), the cross-verse witness.** The mutashābihāt lafẓiyyah are **formula-shaped by nature** — the fourth witness is the one most exposed to the F1/F15 class. Amend to:
> *(4) … cross-verse witness — **subject to the same uniqueness, anchored-run and contained-formula shields as the ḥadīth witness, since repeated-phrase similarity is this witness's normal condition rather than its exception.***

**AMENDMENT 4 — the genre clause (from F4).** Add:
> *A witness is weighed as the genre it is. Prophetic narration, imāmī khabar, khuṭba, treatise and **rijāl entry** are distinct genres, declared per source; a rijāl work witnesses to **narrators**, never to a verse.*

> **MINORITY (2 seats, S1, S2):** approve as written — *"the doctrine is a constitution, not a manual; amendments 1 and 4 belong in the engine spec."* **Recorded.** The majority answers that a doctrine which asserts a false parenthesis about 22% of the ocean is not yet a constitution.

## Q4 — the precedence doctrine when sīrah and graded ḥadīth diverge

### 🗳 BENCH VOTE: **the layer carries the TRANSMITTED POSITIONS BELOW, in full, with attribution — and issues no ruling of its own. 30 for · 0 against.**

**The governing principle the bench affirms first:** the waqf records that the traditions *have* precedence doctrines; it never adopts one. Where the schools differ, **the divergence is the deliverable.** Any engine that silently prefers one witness has adjudicated.

**The positions the layer MUST carry (Sunnī side) — ⚠️ ESTIMATE as to exact wording; each must be re-sourced verbatim before publication, per the grade-verbatim law:**

| # | Transmitted position | Carried as | Must also carry |
|---|---|---|---|
| 1 | The muḥaddithūn's primacy of the graded isnād over maghāzī/sīrah akhbār | the majority ḥadīth-critical position | that it is a statement about **isnād rigour**, not about falsity |
| 2 | The report from **Aḥmad b. Ḥanbal** on maghāzī / malāḥim / tafsīr being subject to laxity (*"laysa lahā aṣl"*) | verbatim, attributed, **with its chain and its disputed reading** | the counter-reading: laxity of *isnād*, not licence to fabricate |
| 3 | **Ibn Isḥāq's** contested standing — commended by al-Zuhrī and Shuʿba; criticised by Mālik and Hishām b. ʿUrwa; used by al-Bukhārī only in support | **all positions side by side, none resolved** | that "criticised" and "abandoned" are different verdicts |
| 4 | **Al-Wāqidī** — declared weak/matrūk by the jumhūr of ḥadīth critics, yet retained by historians for chronology | the **riwāya / tārīkh split**: rejected for narration, consulted for sequence | the split itself is the doctrine, not a compromise |
| 5 | The **tasāhul** convention — a lighter standard outside aḥkām (faḍāʾil, sīrah, raqāʾiq) | attributed to its named proponents, **with the dissent that rejects tasāhul entirely** | that this convention is itself disputed |

**The positions the layer MUST carry (Shīʿī side):**

| # | Transmitted position | Carried as | Note |
|---|---|---|---|
| 6 | The **Akhbārī / Uṣūlī** divergence on the authority of the Four Books' akhbār | two standing schools, both live | this single axis reframes every Shīʿī grade in the bank |
| 7 | The rijāl authorities — **al-Najāshī**, **al-Ṭūsī** (*Fihrist* / *Rijāl*), **Ibn al-Ghaḍāʾirī** | named evaluators, verdicts verbatim | **Ibn al-Ghaḍāʾirī is already held — 226 entries (F4)**, and *his own attribution is itself contested*, which must be carried too |
| 8 | **al-Majlisī's** *Mirʾāt al-ʿUqūl* classifications vs **Behbudī's** stricter *Ṣaḥīḥ al-Kāfī* | **two graders who disagree on the same rows** | this is the Shīʿī T2b, and the bank already holds both (F3) |
| 9 | The *maghāzī*/*tārīkh* material's status relative to graded akhbār | the parallel to (1) | must not be assumed identical to the Sunnī position |

**The mechanical rule the bench proposes, and the only one it will endorse:**
> **THE DIVERGENCE RECORD.** Where sīrah and graded ḥadīth diverge, the layer emits a **conflict node** carrying: both reports verbatim · both carriers named · each one's transmitted grade or its declared absence · **every precedence doctrine from the table above that a school would apply to this pair, attributed** · and **no verdict field**. The node is complete when a reader can see *who would decide it which way and on what stated ground* — and it is defective the moment it contains a preference the waqf itself supplied.

**Two cautions this bench is uniquely placed to give:**
1. **Do not let precedence become a silent sort order.** If the UI ranks graded ḥadīth above sīrah by default, the waqf has adopted position (1) without voting on it. **Rank by nothing; present both; let the doctrine labels do the work.**
2. **A grade is not a precedence.** F16 shows the grades are modern. Using a 20th-century *ṣaḥīḥ* to outrank a 2nd-century *sīrah* report is a real methodological claim that some schools reject outright — it must be visible as a claim, never applied as plumbing.

## Q5 — growing the ocean from the measured 68,921 toward 120,000

### 🗳 BENCH VOTE: **four acquisitions, in the order below, reaching ≈119,172 — 27 for · 3 dissenting on order.**

**The ordering principle the bench adopted (and asks the council to adopt):** ① **complete a canon before broadening one** · ② **proven pipeline before new pipeline** · ③ **score every acquisition on Δ-verses-covered, not Δ-rows** (gap-highlight #8).

| Order | Acquisition | Rows ⚠️ ESTIMATE | Running total | Why this position |
|---|---|---|---|---|
| **1** | **Tahdhīb al-Aḥkām** (al-Ṭūsī) | ~13,590 | 82,511 | Completes the Shīʿī canon. **Same pipeline** — `thaqalayn-api.net` already built the 32 held books (F6). Highest doctrinal value per row in the whole list: without it the waqf ships a Shīʿī bank missing half its canon. |
| **2** | **al-Istibṣār** (al-Ṭūsī) | ~5,511 | 88,022 | **The Four Books are then complete.** Ships with (1) — same author, same source, one ingest. |
| **3** | **Musnad Aḥmad b. Ḥanbal** | ~27,647 | 115,669 | The largest single Sunnī gain available, and the one most likely to move **Δ-verses-covered** — a musnad's spread is topical, not legal, so it should reach verses the sittah's fiqh-ordered books miss. |
| **4** | **Sunan al-Dārimī** | ~3,503 | **119,172** | Cheap, canonical-adjacent, and with (3) it completes the **"nine books."** Lands **0.7% short of 120,000** — the bench recommends stopping here rather than overshooting for a round number. |

**Reserve tier, if the council wants headroom past 120,000:** Ṣaḥīḥ Ibn Ḥibbān (~7,491) · Ṣaḥīḥ Ibn Khuzayma (~3,079) · al-Mustadrak (~8,803) · Wasāʾil al-Shīʿa (~35,850) · Mustadrak al-Wasāʾil (~23,000) · al-Sunan al-Kubrā, Bayhaqī (~21,812) · the two Muṣannafs (~59,000 combined) · Biḥār al-Anwār (very large).

**Five conditions the bench attaches to any acquisition — these are not optional:**
1. **sha256 + licence sealed at ingest**, and — closing gap-highlight #7 — **the source URL and retrieval date recorded per book**. The corpus's own isnād is currently undocumented; do not extend an undocumented chain.
2. **Genre declared at ingest** (F4). Never repeat the Ghaḍāʾirī error in a 27,000-row book.
3. **Grade field mapped or its absence declared** (law 9). ⚠️ **Tahdhīb and Istibṣār carry no standard per-report grade field** — expect them to land as T6, and say so *before* ingest so the tier ladder's shape change is predicted, not discovered.
4. **Schema-probed before any iterator** (law 2). Musnad Aḥmad's musnad ordering is *by Companion*, not by book/chapter — the sunnī `{book, hadith_in_book}` shape will not fit unchanged.
5. **Score published per acquisition:** Δ-rows **and** Δ-verses-covered **and** Δ-narration-families (F9). An acquisition that adds 27,000 rows and 40 verses is a different purchase from one that adds 3,500 rows and 400 verses, and today the waqf could not tell them apart.

> **DISSENT — Seats H1, H5 (Shīʿī) and C3 (comparative):** put **Musnad Aḥmad first**. Reason: it is the single largest gain and the Sunnī side carries the CROWN tier, so ocean growth compounds fastest there. **Recorded, not adopted** — the majority holds that *completeness of a canon* outranks *size of a gain*, and that shipping a Shīʿī bank without Ṭūsī's two books is a coverage-honesty defect (F6) that no amount of Sunnī growth repairs.

---

## THE BENCH'S CLOSING NOTE

The engines built tonight are **better than the prose that describes them**. E2, E5 and E7 held every mechanical test this bench could devise — including a full sweep of all 2,392 full-quote pairs against their own ≥90% law, which passed without a single violation. Twice the bench raised an alarm and twice **the engine was right and the checker was wrong** (skill law 18, earning its place again).

Every defect that survived is in the **narration layer** — the prose, the gap lists, the headline table: a verse that is a substring of the basmalah, a count of recitations presented as a count of explanations, two graders unnamed, a rijāl book counted as ḥadīth, an absence declared over an asset already held. **Five sentences, four files.** That is the whole distance between this audit and one that cannot be attacked.

**The one thing the bench would say if it could say only one thing:** the waqf has measured *presence* to a very high standard and has not yet measured *use* — and 92% of the time, the answer to "why is this verse in this ḥadīth" is "because he recited it in prayer." **Build the `use` classifier before building anything else in P32.**

*Machine-scholar bench. Human review invited, never waited for. Nothing in this repository was edited.*