# COUNCIL OF 120 — BENCH 1 · ḤADĪTH SCIENCES (30 seats)
*ʿulūm al-ḥadīth · muṣṭalaḥ · rijāl ×4 · ʿilal · gharīb al-ḥadīth · nāsikh wa-mansūkh of ḥadīth · mukhtalif al-ḥadīth · takhrīj · tadlīs · ziyādat al-thiqa · the 20 knowledge areas*

**MACHINE-SCHOLAR COUNCIL — labeled as such. Human review invited, never waited for.**
**REVIEW AND COUNSEL ONLY — this bench edited nothing in the corpus or the layers.**
Sitting: 2026-09-04, after E1–E7 landed (all files ≥1,200 s quiescent at read time ✅ VERIFIED).
Scope: the verified link layer `05-READINGS/work/_hadith_VERIFIED.json` and the quality-gate question.

---

## VERDICT

**THE LINK LAYER IS SOUND IN SUBSTANCE AND OVERSTATED IN ITS WARRANT.** Four-fifths of its
22,188 links carry real, meaning-bearing commentary material — the bench replicated the
ocean census exactly and found no fabrication. But its "verification gate" **rejected 0 of
22,188 links** because it re-derives with the *same* 12-word-shingle method that built the
selection, and that shingle **does not distinguish isnād from matn**: **3,662 links (16.5%)
are established by a transmission chain alone**, where the commentator was quoting a
*different* report that shares the chain. The FORMULAIC-SHINGLE and UNIQUENESS laws (7, 16,
17) — written for quote-detection against the Qurʾān — were **never applied to the ḥadīth
link layer**. That is this bench's principal finding and the commission's principal miss.
Second: the per-āyah cap of 9 is real (828 verses pinned at exactly 9), and **the uncapped
pool already exists on disk** at 42,214 links — the remedy is a re-run, not a rebuild.
Third: the grade surface is honest but **entirely modern on the Sunnī side (nine graders,
not one classical)**, and tier T5 mis-attributes 864 non-Majlisī verdicts to Mirʾāt al-ʿUqūl.

**MAXIMUM DEFENSIBLE GATE ON HELD DATA: a 4-tier ladder topping out at 9,675 links** (§ F3).
Anything stronger requires the rijāl acquisition and would be *invented, not transmitted*.

---

## 1 · WHAT THE BENCH INDEPENDENTLY REPLICATED (the all-clear, stated first)

A false all-clear and a false alarm cost the same trust. These numbers were re-derived by
this bench from `00-CORPUS` **without reading E1 first**, then compared:

| Claim (E1 / THE-ANSWERS) | Bench's independent measure | Verdict |
|---|---|---|
| Ocean 68,921 rows | 68,921 | ✅ VERIFIED — exact |
| Sunnī 36,390 in 7 books | 36,390 / 7 | ✅ VERIFIED |
| Shīʿī 32,531 in 32 books | 32,531 / 32 | ✅ VERIFIED |
| Sunnī rows w/ grade field 21,185 | 21,185 (58.2%) | ✅ VERIFIED |
| Shīʿī rows w/ grade field 14,805 | 14,805 (45.5%) | ✅ VERIFIED |
| Links 22,188 on 4,603 verses | 22,188 / 4,603 | ✅ VERIFIED |
| Per-āyah max 9 · median 4 · mean 4.82 | 9 / 4 / 4.82 | ✅ VERIFIED |
| Grade-class census 10,382 / 9,057 / 2,749 | identical | ✅ VERIFIED |
| Tier ladder partitions the ocean exactly | 15,152+16,024+2,628+2,533+53 = 36,390 ✅; 14,805+17,726 = 32,531 ✅ | ✅ VERIFIED — arithmetically closed |

**The engines are not lying.** Every finding below is about *warrant and labelling*, not
about invented data. Two further exonerations the bench specifically hunted for and did
**not** find:

- **F0-a · The "verse-echo" hypothesis is REFUTED.** The bench suspected the gate's shared
  12-word run might simply be the āyah itself, quoted by both ḥadīth and tafsīr. Tested by
  substring containment against the āyah **and its ±2 neighbours** (a report quoting
  scripture rarely stops on the āyah boundary): **only 8 of 2,923 sampled links (0.3%)** are
  verse-echo-only. 99.7% carry non-scriptural text. `CH1_probeB.json` ✅ VERIFIED.
  *(The bench's first pass of this probe used run-set intersection, which is structurally
  void because most āyāt are shorter than 12 words. Law 18 — check the checker — caught it;
  the corrected substring test is the one reported. The first pass is discarded, not cited.)*
- **F0-b · GRADE-VERBATIM (law 9) is NOT violated by the integer `grade` field.** The bench
  suspected the collapsed `grade: 0|1|2` was a fabricated grading. It is **only a sort
  tiebreaker** (`hadith_layer.py:585`). What renders on the page is `grading()`
  (`hadith_layer.py:516-526`), which reads the corpus's own `grades` array verbatim with the
  grader named, and whose own docstring says *"Never a grade this file invented, and never
  the sort key dressed up as a scholar's verdict."* **Charge withdrawn.** ✅ VERIFIED

---

## 2 · FINDINGS

### 🔴 F1 · THE ISNĀD-SHINGLE CLASS — 3,662 links (16.5%) rest on a transmission chain alone
**SEVERITY: HIGH** · `01-TOOLKIT/verify_hadith_links.py:72,115` · `CH1_full_census.json`

The gate at line 115 accepts a link if **any** 12-word run is shared between the ḥadīth text
and the commentary cell. Line 72 feeds it `r.get("matn") or r.get("ar")` — **which contains
the isnād**. A chain such as `عن سعيد بن جبير عن ابن عباس` transmits *hundreds* of distinct
reports. When Ibn Kathīr quotes any one of them under an āyah, the gate "proves" he cited
**this** one.

Found by law 16's own instrument — *always read a top hit*. The bench opened the actual
matched runs (`ch1_whatmatched.py`). Verbatim:

| āyah | link | the run the gate matched on | reading |
|---|---|---|---|
| 94:3 | Tirmidhī 878 (Ibn Kathīr) | `السائب عن سعيد بن جبير عن ابن عباس قال قال رسول الله` | **pure isnād** |
| 22:54 | Nasāʾī 1952 (Ibn Kathīr) | `شعبه عن ابي بشر عن سعيد بن جبير عن ابن عباس قال` | **pure isnād** |
| 26:4 | al-Kāfī ii.1144 (Nūr al-Thaqalayn) | `علي بن ابراهيم عن ابيه عن القاسم بن محمد عن المنقري عن` | **pure isnād** |
| 89:5 | Man Lā Yaḥḍuruh ii.1298 (ʿAyyāshī) | `عبد الله بن سنان عن ابي عبد الله عليه السلام قال ان` | **pure isnād** |
| 5:39 | Muslim 4410 (Ibn Kathīr, 42 runs) | `اذا سرق فيهم الشريف تركوه واذا سرق فيهم الضعيف اقاموا عليه الحد` | **genuine matn** ✅ |
| 3:6 | Muslim 6722 (Baghawī) | `الكتاب فيعمل بعمل اهل النار فيدخلها وان احدكم ليعمل بعمل اهل النار` | **genuine matn** ✅ |

94:3 is the decisive case. The āyah is *"which had weighed upon your back"*; the linked report
is *"the Black Stone descended from Paradise whiter than milk."* Nothing connects them but
Saʿīd b. Jubayr ← Ibn ʿAbbās.

**Measured over the full population** (all 22,188 links, not a sample), with two independent
instruments — document frequency of each shared run across all 68,921 reports (law 7's own
instrument) and an isnād-particle density test, both validated against known positives
(law 5) before use:

| class | meaning | links | share |
|---|---|---|---|
| **A** | unique gram **+** non-isnād text — attribution evidentiary | 10,625 | 47.9% |
| **B** | chain **unique to this report** — a legitimate isnād fingerprint | 2,971 | 13.4% |
| **C** | matn shared with parallel narrations — *content right, which copy ambiguous* | 4,930 | 22.2% |
| **D** | **isnād / boilerplate only — attribution NOT established** | **3,662** | **16.5%** |

Sunnī D = 3,253 (16.4%) · Shīʿī D = 409 (17.4%) — **the defect is tradition-neutral**, which
is itself worth recording. **87 verses have no link of any other class**: every report shown
there is attribution-unestablished.

Class **C is not a defect** — it is law 8 (*ambiguity is data*): the matn is correct, only the
collection attribution is soft (Bukhārī and Muslim carrying the same matn). It must be
*labelled*, not dropped.

> **The bench does not ask for these links to be deleted.** Class D links are still *reports a
> mufassir placed near this āyah*. What must change is the **claim**: "cited by al-Ṭabarī on
> this āyah" is not warranted for class D; "in the vicinity of al-Ṭabarī's comment, chain-matched
> only" is.

### 🔴 F2 · THE VERIFICATION GATE REJECTED NOTHING — it is a restatement, not a check
**SEVERITY: HIGH** · `01-TOOLKIT/verify_hadith_links.py` · SINGLE-SOURCE LAW (11)

```
_hadith_FINAL.json    4,603 verses  22,188 links
_hadith_VERIFIED.json 4,603 verses  22,188 links
links dropped by the gate : 0
links with eds TRIMMED    : 0
links untouched           : 22,188      ✅ VERIFIED (ch1_full.py diff)
```

The file's own docstring is admirable — *"an index is a claim about the data, not the data…
A page that says 'cited by al-Ṭabarī' when al-Ṭabarī said no such thing is a fabricated
attribution to a scholar dead eleven hundred years."* The bench endorses every word. But the
gate **re-derives the link with the same 12-gram shingle method that built the selection**,
so it cannot fail. A 100% pass rate on 22,188 independent claims is not a clean bill of
health; it is the signature of a tautology. The three rejection counters
(`hadith record missing`, `match NOT re-derivable`, `some editions dropped`) all read zero.

**This is the check that F1 slipped through.** A genuinely independent gate must vary the
method — matn-only text (isnād stripped), or a uniqueness requirement, or both.

### 🟠 F3 · THE CAP OF 9 — and the uncapped pool is **already on disk**
**SEVERITY: MEDIUM-HIGH** · `05-READINGS/work/` · ✅ VERIFIED

The cap is real and its signature is unmistakable — the per-āyah histogram falls
monotonically 793 → 597 → 481 → 446 → 364 → 348 → 354 → 392 and then **jumps to 828 at
exactly 9**. 828 verses (18.0%) are pinned at the ceiling.

E1 reported the cap. **What the commission did not report is that the pre-cap pool survives:**

| file | verses | links | max/āyah | median |
|---|---|---|---|---|
| `_hadith_links_BOTH.json` (**uncapped**) | 4,905 | **42,214** | **128** | 5 |
| `_hadith_selected9.json` (cap applied; name says it) | 4,905 | 25,653 | 9 | 5 |
| `_hadith_FINAL.json` → `_hadith_VERIFIED.json` (shipped) | 4,603 | 22,188 | 9 | 4 |

**20,026 links (47.4% of what was discovered) and 302 whole verses were discarded** by the
cap + dedup + reuse-filter. 1,421 verses in the uncapped pool carry more than 9. The rebuild
Q-4 asks about is therefore **a re-run of an existing artifact, not a new engine** — an
important and favourable fact for costing P32.

### 🟠 F4 · TIER T5 MIS-ATTRIBUTES 864 verdicts to Mirʾāt al-ʿUqūl
**SEVERITY: MEDIUM** · `E5_tiers_union.json` → `THE-ANSWERS.md:28` · ✅ VERIFIED

E5 names the tier **"T5 shia graded (Mirʾāt al-ʿUqūl verbatim) — 14,805"**, and THE-ANSWERS
repeats *"Majlisī-graded (Mirʾāt al-ʿUqūl, held verbatim on al-Kāfī) — 14,805 rows."*

Measured, by grader:

| grader | grade strings | which work |
|---|---|---|
| al-Majlisī | 13,941 | Mirʾāt al-ʿUqūl |
| **Behbūdī** | **1,966** | **Ṣaḥīḥ al-Kāfī — a modern reduction, not Majlisī** |
| **Moḥseni** | **825** | **Muʿjam al-Aḥādīth al-Muʿtabara — not Majlisī** |

And the 14,805 rows are **not all al-Kāfī**: al-Kāfī's 8 volumes supply 13,981; the other 824
come from al-Khiṣāl (4), Man Lā Yaḥḍuruh vol. 5 (390), Muʿjam al-Muʿtabara (428), ʿUyūn vol. 1
(2). By the layer's own standard — *"a fabricated attribution to a scholar"* — putting
Behbūdī's and Moḥseni's verdicts under Majlisī's name is exactly the error the pipeline was
built to prevent. **Fix: rename T5 "Shīʿī graded (grader named per row)" and split it.**

### 🟠 F5 · T5 IS NOT A QUALITY TIER — 43% of it is a *weak* verdict
**SEVERITY: MEDIUM** · `CH1_full_census.json` · ✅ VERIFIED

The Sunnī side of the ladder carefully separates acceptable (T2, 16,024) from ḍaʿīf (T3,
2,628) from graders-differ (T2b, 2,533). **The Shīʿī side gets no such split** — all 14,805
graded rows sit in one tier that reads, in a ladder, as parallel to a ṣaḥīḥ/ḥasan tier.
Measured split of those 14,805 rows:

| class | rows | share |
|---|---|---|
| weak family (ḍaʿīf / majhūl / mursal / marfūʿ) | **6,365** | **43.0%** |
| acceptable family (ṣaḥīḥ / ḥasan / muwaththaq / muʿtabar) | 6,078 | 41.1% |
| `لم يخرجه` + weak | 1,256 | 8.5% |
| mixed wording (`حسن كالصحيح`, `ضعيف على المشهور`) | 386 | 2.6% |
| `لم يخرجه` + acceptable | 347 | 2.3% |
| other / unparsed | 178 | 1.2% |

Two sub-findings the bench flags for the reading layer:

- **`لم يخرجه` (1,734 strings, Behbūdī) is not a grade.** It records that Behbūdī *did not
  include* the report in his Ṣaḥīḥ al-Kāfī. Carrying it in a "grade" column is a category
  error; it is an *omission notice*, transmitted as such.
- **`مجهول` (2,277) / `مرسل` (764) / `مرفوع` (268) are isnād-form or rijāl descriptors**, not
  authenticity verdicts. A tier gate that reads them as grades will mis-sort them.

### 🟠 F6 · THE STRONGEST GATE ON OFFER (31,176) CONTAINS **ZERO** SHĪʿĪ REPORTS
**SEVERITY: MEDIUM-HIGH** — a doctrine conflict, not a data error · `THE-ANSWERS.md:27`

THE-ANSWERS' "Ṣaḥīḥ-tier ocean max = T1 + T2 = 31,176" is arithmetically correct
(15,152 + 16,024) and the two sets are properly disjoint (Bukhārī/Muslim carry no grade
field, so cannot enter T2) ✅ VERIFIED. But T1 and T2 are **both Sunnī-only by construction**:
T1 is the Ṣaḥīḥayn; T2 is built from the Sunnī grader vocabulary. The waqf's top quality gate
is therefore **100% Sunnī** — against `hadith_layer.py:36-42`'s own governing law: *"Sunni
and Shi'i reports interleave in a single list… never grouped, coloured, badged or sectioned
by tradition… the hadith layer is not the place this project starts dividing the ummah."*

**This is fixable today from held data.** Applying one symmetric rule to both corpora
(acceptable = every named grader on the row calls it acceptable; Ṣaḥīḥayn = collection is the
claim):

| tier (symmetric) | total | Sunnī | Shīʿī |
|---|---|---|---|
| T1 Ṣaḥīḥayn (collection is the claim) | 15,152 | 15,152 | 0 |
| T2 all named graders acceptable | **21,944** | 15,866 | **6,078** |
| T2b graders differ (preserved) | 3,101 | 2,690 | 411 |
| T3 no grader calls it acceptable (kept visible) | 10,945 | 2,629 | 8,316 |
| T6 ungraded in held data | 17,779 | 53 | 17,726 |
| **TOTAL** | **68,921** | 36,390 | 32,531 |

Sum closes exactly on the measured ocean ✅ VERIFIED. **A symmetric top gate is 37,096, not
31,176**, and it includes 6,078 Shīʿī reports. *(The bench's T2 rule is marginally stricter
than E5's — a single dissenting grader sends a row to T2b — which is why Sunnī T2 reads
15,866 here vs 16,024 there. Two methods, two numbers, reported side by side and never
blended.)*

### 🟡 F7 · EVERY SUNNĪ GRADER IN THE HELD CORPUS IS MODERN — not one classical authority
**SEVERITY: MEDIUM** · `CH1_grade_surface.json` · ✅ VERIFIED

All nine graders carrying the Sunnī grade surface died in, or are living in, the 20th–21st
century:

| grader | rows | d. |
|---|---|---|
| al-Albānī | 19,081 | 1420/1999 |
| Zubair ʿAlī Zaʾī | 19,033 | 1435/2013 |
| Shuʿayb al-Arnaʾūṭ | 6,413 | 1438/2016 |
| ʿAbd al-Fattāḥ Abū Ghudda | 5,692 | 1417/1997 |
| M. Muḥyī al-Dīn ʿAbd al-Ḥamīd | 5,171 | 1392/1972 |
| M. Fuʾād ʿAbd al-Bāqī | 4,314 | 1388/1968 |
| Aḥmad Muḥammad Shākir | 3,729 | 1377/1958 |
| Bashshār ʿAwwād Maʿrūf | 2,425 | living |
| Sālim al-Hilālī | 1,858 | living |

E1's honest_gaps correctly declares that **rijāl** is unheld. It does **not** declare that the
**grades themselves are an exclusively modern layer**. No Ibn Ḥajar, no al-Dhahabī, no
al-Dāraquṭnī's *ʿIlal*, no al-Mizzī, no Ibn Abī Ḥātim. Same on the Shīʿī side: al-Majlisī
(d. 1110/1699) is the earliest, and Behbūdī and Moḥseni are contemporary. For a waqf whose
watchword is *transmitted, never invented*, the provenance of the grade layer must be stated
on the page: **these are muḥaqqiq judgements of the last century, not the classical ḥukm.**
1,671 distinct Sunnī grade strings and 352 Shīʿī ones are carried verbatim — that part is
exemplary.

### 🟡 F8 · "the OTHER 24 Shīʿī books carry no grade field" — measured, it is **20**
**SEVERITY: LOW (accuracy)** · `E1_ocean_census.json:143` · ✅ VERIFIED

32 Shīʿī books total. **12** carry at least one grade (al-Kāfī ×8, al-Khiṣāl, Man Lā Yaḥḍuruh
vol. 5, Muʿjam al-Muʿtabara, ʿUyūn vol. 1) → **20 carry none**. The figure 24 appears to come
from 32 − 8 (al-Kāfī's volumes), which overlooks the four non-Kāfī books that *are* graded.
Offered under I-3 for correction in the record.

### 🟡 F9 · TAFSĪR-RELEVANCE — the bench's estimate, and why this layer beats the quote census
**SEVERITY: informational — but decisive for Q-4** · ⚠️ ESTIMATE, stated as such

The bench read 12 links end-to-end (8 Sunnī, 4 Shīʿī, seeded sample, `ch1_sample12.py`) and
judged each for *bearing on the āyah's meaning* versus *merely reciting it*:

| verdict | n | examples |
|---|---|---|
| **directly meaning-bearing** (classical tafsīr material) | 3 | 3:6 ← Muslim's 40+40+40 womb-formation report — the standard gloss on *yuṣawwirukum fī l-arḥām*; 5:39 ← Muslim's Makhzūmī theft report under the repentance-after-theft verse; 16:44 ← Bukhārī on waḥy as the miracle given |
| **meaning-bearing at passage level** (filed 1–3 āyāt off) | 3 | 8:66 ← Abū Dāwūd on the Badr ransom (the report recites 8:67-68); 4:66 ← Bukhārī, matter of 4:69; 38:18 ← Muslim on Dāwūd's fast and prayer |
| **not evidently meaning-bearing** | 6 | 94:3, 22:54, 26:4, 6:27, 89:5, 11:123 — **5 of these 6 are class D or isnād-matched** |

**Estimate of the tafsīr-relevant share: 60–65%** carry a meaning-bearing relation *with sound
attribution* (≈ classes A+B = 13,596 links, 61.3% — the mechanical proxy and the qualitative
read agree, which is why the bench offers the range); a further ~22% (class C) are
content-correct with ambiguous collection attribution; ~16.5% (class D) are not established.
Read at 12 links the qualitative estimate carries a wide interval and is **one bench's
judgement, not a measurement** — it is labelled ESTIMATE for that reason.

**The structural point matters more than the number.** These links are **tafsīr-mediated**: a
named mufassir chose to place this report under this āyah. That is a far stronger
relevance signal than a quote-match. E2's 9,962 uncapped quote-pairs are the *opposite*
relation — a ḥadīth that **recites** the verse, which is precisely the "merely quotes it"
class. **The two layers measure different things and the verified layer is the more
tafsīr-relevant of the two.** This governs the bench's Q-4 vote.

---

## 3 · MECHANICAL TABLE

| # | claim under audit | measured | source | label | severity |
|---|---|---|---|---|---|
| 1 | ocean rows | 68,921 (S 36,390 / Sh 32,531) | bench re-derived; = E1 | ✅ VERIFIED | — |
| 2 | verified links / verses | 22,188 / 4,603 | `_hadith_VERIFIED.json` | ✅ VERIFIED | — |
| 3 | per-āyah cap | max 9; **828 verses pinned at 9** (18.0%) | bench | ✅ VERIFIED | 🟠 |
| 4 | uncapped pool exists on disk | **42,214 links / 4,905 verses / max 128** | `_hadith_links_BOTH.json` | ✅ VERIFIED | 🟠 |
| 5 | discarded by cap+dedup | **20,026 links (47.4%), 302 verses** | bench diff | 📊 CALCULATED | 🟠 |
| 6 | gate rejections | **0 dropped, 0 eds trimmed, 22,188 untouched** | bench diff FINAL↔VERIFIED | ✅ VERIFIED | 🔴 |
| 7 | attribution class A (evidentiary) | 10,625 (47.9%) | `CH1_full_census.json` | 📊 CALCULATED | — |
| 8 | class B (unique-chain fingerprint) | 2,971 (13.4%) | " | 📊 CALCULATED | — |
| 9 | class C (parallel matn, copy ambiguous) | 4,930 (22.2%) | " | 📊 CALCULATED | 🟡 |
| 10 | **class D (isnād/boilerplate only)** | **3,662 (16.5%)**; 87 verses all-D | " | 📊 CALCULATED | 🔴 |
| 11 | verse-echo-only links | 8 / 2,923 = **0.3%** — hypothesis refuted | `CH1_probeB.json` | ✅ VERIFIED | ✅ clear |
| 12 | grade-verbatim law honoured on the page | yes — `hadith_layer.py:516-526` | code read | ✅ VERIFIED | ✅ clear |
| 13 | Sunnī graded rows | 21,185 / 36,390 (58.2%); Ṣaḥīḥayn 15,152 have no grade field | bench | ✅ VERIFIED | — |
| 14 | Sunnī graders — all modern | 9 graders, earliest d. 1958 | `CH1_grade_surface.json` | ✅ VERIFIED | 🟡 |
| 15 | Shīʿī graders | Majlisī 13,941 · Behbūdī 1,966 · Moḥseni 825 | " | ✅ VERIFIED | 🟠 |
| 16 | T5 "Mirʾāt al-ʿUqūl verbatim = 14,805" | **864 verdicts are not Majlisī's**; 824 rows are not al-Kāfī | bench | ✅ VERIFIED | 🟠 |
| 17 | Shīʿī "graded" rows that are **weak** verdicts | 6,365 (43.0%) — no T3 equivalent exists Shīʿī-side | `CH1_full_census.json` | 📊 CALCULATED | 🟠 |
| 18 | `لم يخرجه` carried as a grade | 1,734 strings — an omission notice, not a grade | bench | ✅ VERIFIED | 🟡 |
| 19 | Shīʿī books with zero grades | **20**, not 24 | bench | ✅ VERIFIED | 🟡 |
| 20 | top gate 31,176 = T1+T2 | arithmetic correct, sets disjoint — **but 0 Shīʿī** | E5 + bench | ✅ VERIFIED | 🟠 |
| 21 | symmetric top gate (both traditions) | **37,096** (T1 15,152 + T2 21,944); ladder closes on 68,921 | bench | 📊 CALCULATED | — |
| 22 | GATE-1 (A\|B × T1\|T2) | **9,675 links** | `CH1_gate_matrix.json` | 📊 CALCULATED | — |
| 23 | GATE-2 (A\|B\|C × T1\|T2\|T2b) | 15,696 links | " | 📊 CALCULATED | — |
| 24 | tafsīr-relevant share | **60–65%** sound-attribution meaning-bearing | 12-link read + A+B proxy | ⚠️ ESTIMATE | — |

**Artifacts this bench wrote** (new evidence files, no corpus touched):
`10-WAQF/build/LENS-AUDIT/CH1_probeA.json` · `CH1_probeB.json` · `CH1_isnad_census.json` ·
`CH1_grade_surface.json` · `CH1_full_census.json` · `CH1_gate_matrix.json`

---

## 4 · GAP-HIGHLIGHT DUTY — what the commission itself missed

1. **Laws 7, 16 and 17 were never applied to the ḥadīth link layer.** They were born for
   quote-detection against the Qurʾān (the oath-family false pairs) and the audit applied them
   there brilliantly. The *identical* pathology — a high-document-frequency shingle forging
   links — lives in the link layer as the **isnād**, and nothing checked for it. **The oath
   family was 3,580 false pairs; the isnād family is 3,662 unwarranted links.** Almost the
   same size, found by the same law, missed because the law was filed under "quotes".
2. **The remedy for the cap is already on disk.** E1 reported the cap as a finding; nobody
   ran `ls` on the work directory. `_hadith_links_BOTH.json` has held 42,214 uncapped links
   since 2026-08-26.
3. **The verification gate's 100% pass rate was never treated as a signal.** A check that
   rejects nothing should have triggered law 18 (check the checker) on its first run.
4. **The grade surface's modernity is undeclared.** "Rijāl is not held" is declared; "every
   grade we hold is a 20th–21st-century judgement" is not, and it is the more consequential
   fact for a reader who assumes ṣaḥīḥ means the classical ḥukm.
5. **The top quality gate silently became mono-traditional** — against the layer's own
   most emphatic law. No engine noticed because no engine tests tier composition by tradition.
6. **The 120,000 target is a ROW count, not a distinct-report count.** Wasāʾil al-Shīʿa
   (~35,000) and Biḥār al-Anwār are *compilations that re-transmit the Four Books*. Acquiring
   them reaches 120,000 on paper while adding far fewer distinct reports. **The bench asks
   that the target be re-defined on distinct matn**, or the headline number becomes a
   duplication artifact. (The layer already has the instrument: `dupes` is populated on 1,123
   links, so cross-collection identity is modelled — it is simply not applied to the ocean count.)
7. **E2 and the verified layer measure different relations** (recitation vs. tafsīr-citation)
   and THE-ANSWERS' union bracket places them on one axis. The bracket is honestly labelled a
   bracket, so this is not an error — but a reader will take "258 on 112:1" as 258 *explanations*,
   and they are overwhelmingly *recitations of the verse in other contexts*. **The union needs
   a relation-type column, not just a count.**
8. **`dupes` is under-exploited.** 1,123 links carry duplicate-report pointers. That is exactly
   the instrument that would resolve class C (22.2%) from "which copy?" to a named set.

---

## 5 · THE MAXIMUM DEFENSIBLE QUALITY GATE ON HELD DATA (the bench's ruling)

The bench was asked to rule honestly on what gate is buildable **with only held data**. The
answer is a **four-tier ladder on two independent axes** — *attribution strength* (does the
link hold?) and *transmitted grade* (what did named graders say?). Neither axis alone is a
gate; the corpus supports the product of the two.

`CH1_gate_matrix.json` — attribution class × source tier, all 22,188 links:

| source tier | A | B | C | D | total |
|---|---|---|---|---|---|
| T1 Ṣaḥīḥayn | 4,621 | 500 | 1,944 | 900 | 7,965 |
| T2 all graders acceptable | 3,265 | 1,289 | 2,432 | 1,913 | 8,899 |
| T2b graders differ | 1,034 | 372 | 239 | 387 | 2,032 |
| T3 none call it acceptable | 954 | 644 | 213 | 389 | 2,200 |
| T6 ungraded in held data | 751 | 166 | 102 | 73 | 1,092 |
| **total** | **10,625** | **2,971** | **4,930** | **3,662** | **22,188** |

**THE LADDER — four named tiers, nothing invented:**

| tier | rule | links | what may be claimed |
|---|---|---|---|
| **ǪĀṬIʿ** (decisive) | class A × T1 — Ṣaḥīḥayn, unique non-isnād matn reproduced by the mufassir | **4,621** | "al-Ṭabarī/Ibn Kathīr cites this report here" — full attribution |
| **THABIT** (established) | class A\|B × T1\|T2 | **9,675** | attribution established + every named grader acceptable |
| **MAQBŪL** (admissible) | class A\|B\|C × T1\|T2\|T2b | **15,696** | content sound; collection attribution may be a parallel — *labelled* |
| **MAWJŪD** (present, shown, not claimed) | everything else, incl. all class D and T3 | 6,492 | "found in the vicinity of this comment" — **never** "cited by" |

**T3 stays visible** (2,200 links) — *hiding a ḍaʿīf verdict is as dishonest as hiding a ṣaḥīḥ
one*, and E5 already says so. The tier is a **label**, not a filter.

**What is NOT buildable, and the bench refuses to fake it:** anything per-narrator. No
ṭabaqāt, no tadlīs register, no ʿilal, no muʿallal/shādhdh/munkar detection, no isnād-family
clustering with names resolved. `Kitāb al-Ḍuʿafāʾ` of Ibn al-Ghaḍāʾirī (226 rows) is held and
is a *rijāl* work, but 226 rows is a rounding error against 68,921. **The 20-knowledge-areas
gate is not buildable today. E1 says so. The bench confirms it, unanimously.**

---

## COUNSEL — BENCH 1's VOTES

*30 seats. Majority carries; dissents preserved verbatim, never flattened.*

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

**VOTE: (b) — an amended split. 23 for · 6 for (a) · 1 abstain.**

| allocation | (a) as tabled | **bench (b)** | one-line reason |
|---|---|---|---|
| P28 Sīrah Universe 300 → | +1,500 | **+1,200** | E3 is strong and sīrah is a *reading* problem; 1,200 nodes carry the chronology spine, maghāzī registry and sīrah-verse sync without pre-building benches for material not yet acquired |
| **P32** (new pillar) | +1,800 | **+1,600** | **F3: the uncapped rebuild is a re-run of `_hadith_links_BOTH.json`, not a new engine — 200 nodes come off the estimate on measured evidence** |
| **Acquisitions & rijāl reserve** | +300 | **+800** | the binding constraint on every gate above MAQBŪL is *unheld rijāl* (F7); a 300-node reserve cannot absorb Taqrīb-class ingestion + the 20 ungraded Shīʿī books + Musnad Aḥmad |
| **total** | +3,600 | **+3,600** | closes exactly on 30,000 |

**Reasons, one line each:** ① the reserve is the only line that unblocks *all* the others, and
at 300 it is the smallest line funding the largest constraint; ② P32's biggest single
deliverable already exists on disk (F3), which is a measured saving, not an optimistic one;
③ the sīrah expansion is the most speculative of the three because its precedence doctrine
(Q4) is unsettled — funding it hardest, first, buys the least-certain thing.

> **DISSENT (6 seats — the ʿilal and rijāl seats).** *"Take the 300 from P32, not from the
> sīrah. The link rebuild is a re-run; the sīrah spine is not, and E3 measured 2,821 verses
> the held sīrah never touches. Fund the reserve to 800 by cutting P32 to 1,300."*
> **ABSTAIN (1 seat, gharīb al-ḥadīth):** *"No vote until a node is defined in hours. Three
> benches are pricing an unpriced unit."* — the bench records this as a fair objection.

### Q2 · P32 as a NEW pillar vs. an expansion of the Hadith Universe — and its NAME

**VOTE: NEW PILLAR — 27 for · 3 for expansion.**
**NAME: `P32 — AL-SHAWĀHID: THE FOUR-WITNESS PILLAR` — 19 for; `P32 — THE FOUR-WITNESS PILLAR` (as tabled) 8; other 3.**

**Reason:** the Ḥadīth Universe is a **corpus** pillar — it holds reports. P32 holds **relations
between corpora** (link, tier, quote, cross-verse, sīrah join, conflict). Folding a relation
layer into a corpus pillar is what produced F1 in the first place: the link layer was treated
as a property *of* the ḥadīth collection rather than as an evidentiary object with its own
gate. Separating them forces the gate to have an owner.

*On the name:* **shawāhid** (شواهد) is the tradition's own word for corroborating witnesses —
it is precisely the technical term for a supporting report in takhrīj, it carries the
four-witness idea in one indigenous word, and it does not require the reader to already know
this project's internal doctrine. The three "other" votes preferred `al-Qarāʾin` (indicators),
recorded.

### Q3 · The four-witness doctrine wording — approve or amend

**VOTE: APPROVE WITH ONE MANDATORY AMENDMENT — 30 for the amendment (unanimous).**

The doctrine as written (SKILL.md:48-55) is endorsed in full — especially *"the waqf never
adjudicates between witnesses; it makes every conflict VISIBLE with carriers named"* and the
refusal to rank the witnesses. The bench amends one thing:

> **AMENDMENT — THE ATTACHMENT CLAUSE.** *Naming a witness's carrier is not enough; the
> strength of the witness's attachment TO THIS VERSE must be carried with it, in the
> tradition's own vocabulary and by a stated method. A report whose link to an āyah rests on
> a shared transmission chain is not attached to that āyah at all, and it must not appear
> under the same word as a report whose distinctive matn a named mufassir reproduced beneath
> it. Every witness therefore carries two labels, never one: WHO transmits it, and HOW it is
> attached.*

**Reason:** F1 is exactly a four-witness failure. The ḥadīth witness "spoke" on 94:3 with full
carrier naming — Tirmidhī, graded ṣaḥīḥ by four named graders, cited by Ibn Kathīr — and every
one of those attributions is true. The *attachment* is what was false. A doctrine that
validates carriers but not attachment will present F1's 3,662 links as four-witness evidence.

A second, non-blocking amendment carried 21–9 and is recorded rather than pressed:
*"witness (3) SĪRAH and witness (2) ḤADĪTH are not independent — the sīrah is transmitted
through akhbār, often through the same men. Calling them two witnesses risks counting one
witness twice. The doctrine should say so where it says 'itself transmitted through akhbār'."*

### Q4 · The precedence doctrine when sīrah and graded ḥadīth diverge

**VOTE: the layer carries the TRANSMITTED positions below, with attribution, and issues NO
ruling of its own — 30 for (unanimous). The bench refuses to state a precedence of its own,
and asks that the layer be built so that it *cannot*.**

The layer must carry, each as a *named position held by named scholars*, not as the waqf's view:

1. **The muḥaddithūn's isnād-primacy** — the report graded by isnād outranks the ungraded
   khabar. Carried with the remark transmitted from **Aḥmad b. Ḥanbal** that three genres
   *"have no aṣl"* (maghāzī, malāḥim, tafsīr) — **and with the classical dispute over its
   meaning**, since it is read both as a verdict on isnāds and as a remark about genre
   convention. Never carried bare.
2. **Tasāhul vs. tashaddud by subject** — the position (Ibn Ḥanbal, Ibn Mahdī, later
   **Ibn al-Ṣalāḥ** and **al-Nawawī**) that leniency is permitted in faḍāʾil, siyar and
   raqāʾiq, while aḥkām require the strict gate. This is *the* doctrine that governs most
   sīrah/ḥadīth divergence and must be the first thing a reader sees.
3. **The named source-criticism positions on the sīrah transmitters** — carried as disputes,
   with both sides: **Ibn Isḥāq** (Mālik's censure vs. Shuʿba's and al-Zuhrī's praise; Ibn
   Ḥajar's *ṣadūq yudallis*); **al-Wāqidī** (matrūk per Aḥmad, al-Nasāʾī, al-Bukhārī — yet
   relied on by later historians and by Ibn Saʿd); **Sayf b. ʿUmar** (ḍaʿīf, and al-Ṭabarī's
   principal source for the ridda); **al-Kalbī** and **Muqātil** in the tafsīr-sīrah overlap.
4. **Tawātur over āḥād in chronology** — where a sīrah fact is mutawātir or ʿamalī
   (the qibla change, the hijra, the fast of Ramaḍān), it is not displaced by a single āḥād
   report. Transmitted as the uṣūlī position, with the ḥanafī/shāfiʿī differences on
   *mashhūr* noted.
5. **The asbāb al-nuzūl rules of al-Wāḥidī and al-Suyūṭī** — *ṣarīḥ* ("nazalat fī kadhā")
   outranks non-ṣarīḥ ("kunnā naqūlu"); competing asbāb are reconciled by tarjīḥ or by
   taʿaddud al-nuzūl. This is the exact mechanism for the 8:66 case in F9 and it is
   transmitted, not invented.
6. **Shīʿī-side precedence, carried in its own voice and never mapped onto Sunnī categories**:
   al-Kulaynī's own statement in al-Kāfī's preface about his material; the **Akhbārī/Uṣūlī**
   dispute over whether al-Kāfī's reports are accepted wholesale; **al-Majlisī's Mirʾāt** as a
   *later Uṣūlī overlay* on Kulaynī's own claim (this is why F4 matters — the grader must be
   named for the reader to locate the position); ʿarḍ ʿalā l-Qurʾān (measuring a khabar
   against the Book) as a transmitted Shīʿī and Sunnī criterion; **Behbūdī's** modern
   reduction and its contestation.
7. **Where the two traditions' precedence rules themselves diverge, that divergence is the
   record** — displayed, never resolved.

**Mechanical requirement attached to this vote:** every precedence position ships as a
*citable node with its holder named and its opponents named*. If a divergence is displayed
without at least one transmitted position attached to it, the layer has editorialised by
silence — which is the one failure mode this doctrine cannot detect in itself.

### Q5 · Which acquisitions grow the ocean from 68,921 toward 120,000, in what order

**VOTE: the order below — 26 for · 4 dissent on rank 1 vs 2.**
**Precondition carried unanimously: the target is re-defined on DISTINCT REPORTS, not rows (gap 6).**

| # | acquisition | ~rows | why HERE in the order |
|---|---|---|---|
| **1** | **Musnad Aḥmad b. Ḥanbal** | ~27,600 | the single largest honest gain (68,921 → ~96,500) **and** it arrives with **Shuʿayb al-Arnaʾūṭ's** grading apparatus — already the 3rd-largest grader in the corpus, so the tier engine needs no new vocabulary. Best ratio of ocean-growth to gate-growth of any candidate. |
| **2** | **Tahdhīb al-Aḥkām + al-Istibṣār** (al-Ṭūsī) | ~13,600 + ~4,400 | **completes the Shīʿī Four Books** — the direct, measurable fix for F6 and for T6's 17,726 ungraded rows; without these the Shīʿī side of every gate stays structurally thin |
| **3** | **al-Mustadrak** (al-Ḥākim) **with al-Dhahabī's Talkhīṣ** | ~8,800 | the **first classical grader in the building** (F7). Its value is not its size — it is that al-Dhahabī's verdicts break the modern-only monopoly on the grade surface |
| **4** | **Ṣaḥīḥ Ibn Ḥibbān** + **Sunan al-Dārimī** | ~7,500 + ~3,500 | high-authority, well-edited, Arnaʾūṭ apparatus on Ibn Ḥibbān; ~116,000 cumulative |
| **5** | **al-Sunan al-Kubrā** (al-Bayhaqī) | ~21,000 | crosses 120,000 on distinct reports — but heavy overlap with the Six; **acquire only with dedup measured first** |
| **6** | **Wasāʾil al-Shīʿa** | ~35,000 | **deliberately last of the ocean-growers.** It is a *compilation re-transmitting the Four Books*; counting its rows as new ocean is the duplication artifact of gap 6. Acquire for topical access, **not** for the headline count |
| **—** | **Biḥār al-Anwār** | ~ vast | **not recommended for the ocean count at all** — encyclopaedic compilation; ingest as a *reference* layer if at all |

**Run in parallel, on the reserve, and not counted in the ocean** — this is what actually
unblocks the gate ladder above MAQBŪL:

| rijāl acquisition | supplies |
|---|---|
| Ibn Ḥajar, **Taqrīb al-Tahdhīb** (+ Tahdhīb al-Tahdhīb) | per-narrator verdicts, ṭabaqāt — the named blocker in E1's honest_gaps |
| al-Mizzī, **Tahdhīb al-Kamāl** · al-Dhahabī, **al-Kāshif** / **Mīzān al-Iʿtidāl** | the classical layer F7 says is entirely absent |
| al-Dāraquṭnī, **al-ʿIlal** | the only route to ʿilal; without it the ʿilal seats on this bench have nothing to work with |
| Shīʿī: **al-Najāshī**, al-Ṭūsī's **Fihrist** and **Rijāl**, al-Kashshī's **Ikhtiyār**, al-Khūʾī's **Muʿjam Rijāl al-Ḥadīth** | the evaluation tradition for the **20** ungraded Shīʿī books (F8) |

> **DISSENT (4 seats — the rijāl seats).** *"Rank al-Mustadrak with al-Dhahabī FIRST, ahead of
> Musnad Aḥmad. One classical grader changes what the whole 68,921 may be said to be; 27,600
> more rows graded by the same modern hands changes only the size. We are optimising the
> ocean when the gate is the constraint."* The majority answered that Musnad Aḥmad's Arnaʾūṭ
> apparatus is itself a large gate gain and that ranks 1 and 3 are weeks apart, not years —
> **the dissent is recorded because it may be right.**

---

## 6 · THE BENCH'S THREE ASKS OF THE INTEGRATOR

1. **Add the attribution class to the pair registry (E7) before the P32 rebuild is scoped.**
   `CH1_full_census.json` has all 22,188 already classed A/B/C/D. Rebuilding uncapped without
   it multiplies F1 from 3,662 to an estimated ~7,000 links at the 42,214 scale.
2. **Rebuild uncapped from `_hadith_links_BOTH.json` — and do NOT swap in E2's quote census.**
   They measure different relations (F9). The uncapped *tafsīr-citation* layer is the
   meaning-bearing one; E2's quote-pairs are the "merely recites it" class. Both should ship,
   in separate columns, under a **relation-type** label.
3. **Correct three numbers in the record under I-3:** T5's attribution (F4), the "24" ungraded
   Shīʿī books (F8, it is 20), and the Sunnī-only composition of the 31,176 top gate (F6).

---

*Bench 1 of 4 · 30 seats · findings: 9 (2 🔴, 4 🟠, 3 🟡) · 2 charges investigated and
withdrawn (F0-a verse-echo, F0-b grade-verbatim) · 1 checker error self-corrected under law 18
· 6 evidence artifacts written · nothing in the corpus or the layers was edited.*
*A false alarm costs trust exactly as much as a false all-clear — both directions were run.*