# COUNCIL OF 120 — BENCH 2 · THE ARABIC BENCH (ṣarf · naḥw · balāghah · lexicography · rasm/qirāʾāt · ṣīghah/wazn)

*Machine-scholar bench. Thirty charters, reasoned as thirty named specialists; disagreements preserved.
Human scholarly review is invited and never waited for. **Review and counsel only — this bench edited nothing.***

**Sitting:** 2026-09-04, 01:47–02:25 AST · **Scope:** does the sealed 26,400 (P30-SIGHAH + P31-MIZAN) actually give a translator the LANGUAGE witness that the four-witness test requires?
**Corpus sha under audit:** `fbc8744a98b36edb2362ee63a72d77c2079e929b2128dad96b41269c6ba199bb` (identical across all 144 layers read — ✅ VERIFIED, no split-source).
**Quiescence:** all LENS-AUDIT files E1–E7 read at mtimes ≥ 21 minutes old (01:13–01:26 vs read at 01:47). Law 10 honoured.

---

## VERDICT

> **CONDITIONAL PASS WITH THREE HARD DEFECTS AND ELEVEN NAMED GAPS.**
>
> The language witness is **real, honest and unusually well-gated** — the i'rāb gate holds under attack (Attack 3: clean, nothing fakes it), the AWAITED declarations are truthful, and the treebank-derived censuses are genuinely re-derivable. But the layer as sealed is a **CENSUS, not a LOOKUP**: 27 of 144 layers hold no per-verse container at all, only totals plus two exemplars, so a translator standing at a verse cannot query them. Two sealed artifacts (`tvf_v1`, `dossiers_v3`) **contradict each other on 65 verses**, and one balāghah class (`W5-T3a` oath-wāw) is **label-inverted** — the 28 genuine qasam-wāws are filed under "preposition" and 3 comitative wāws are filed under "oath". None of these is fatal; all three are mechanically fixable; none was hidden by intent — but the first two mean the TVF today can tell a translator "a Prophetic ﷺ explanation is transmitted here" and hand him nothing when he asks to read it.
>
> On the coupling question: **the language layers and the ḥadīth links are not merely parallel silos — they use different address spaces and there is no primary key on a transmitted row.** The ṣīghah index addresses `sūrah:āyah:word:seg`; every one of the six JOINS layers and all 22,188 verified links address `sūrah:āyah`. `strata_layer.json` dropped the `id` field that `H1`/`H2` harvests carried. Until a transmitted row has an id, no word-level lens can be built — it would be a name-join, which the Constitution forbids (I-1: *join on IDs, never on names*).

---

## THE BENCH — thirty seats, six chambers

| Chamber | Seats | What it attacked |
|---|---|---|
| **ṢARF** (morphology) | 5 | wazn recoverability per word; VF tags; the ṣamad/aḥad question; passive & jussive force |
| **NAḤW** (syntax) | 5 | the i'rāb gate; iḍāfa candidates; nāʾib al-fāʿil; ḍamīr al-faṣl; the AWAITED-7 |
| **BALĀGHAH** | 5 | qasam, istifhām taqrīrī, taqdīm, iltifāt, ḥadhf, emphasis stacks, bayān terms |
| **LEXICOGRAPHY** | 5 | wujūh wa-naẓāʾir, gharīb, muʿarrab, aḍdād, homograph surface |
| **RASM / QIRĀʾĀT** | 5 | the three-transmission census; word-level variants; `hayta`/`kufuwan`; madd marking |
| **ṢĪGHAH / WAZN** | 5 | the index→layer→TVF chain; addressability; the census-vs-lookup question |

---

## PART I — ATTACK 1: FIVE HARD VERSES WALKED THROUGH THE LAYERS

Method: for each verse, the bench pulled (a) the ṣīghah index per-segment tags via `10-WAQF/framework/sighah_index.py`, (b) every one of the 144 P30/P31 JSON layers that references the verse, (c) the TVF checklist, (d) the dossier. Then it asked one question: **what would a competent translator still have to ask, that NO layer answers?** Each such question is a named gap below.

### 1 · 55:13 — `فَبِأَىِّ ءَالَآءِ رَبِّكُمَا تُكَذِّبَانِ`

**What the layers DO give (credit where due):** the dual is caught three ways — `2D` on `كُمَا`, `تُكَذِّبَ` and `انِ` (index ✅ VERIFIED); `literal_layer` casts "you two / of you two"; TVF fires *"FORM: 2nd-person DUAL address — the rendering must carry 'you two' or a declared device"*; and `fidelity_matrix` **measures** the result: 55:13 = `dropped-or-unclear` in the held English/Urdu editions. That is a genuine, working, four-step chain from Arabic form → translation defect. It is the best thing in the pillar and the bench says so plainly.

**What NO layer answers:**
- **GAP L1 — `أَىِّ` is tagged `INTG` but nothing says the interrogative is *taqrīrī*.** The rhetorical-affirmation reading (the tradition's near-universal reading of this refrain: not a question seeking information but an assertion demanding assent) exists in no layer. `ma_types_layer` classifies *interrogative māʾ* (95 positions); there is **no istifhām-force layer at all** for `أَيّ`, `هَل`, `أ`, `مَن`, `كَيْف`. A translator rendering "So which of the favours…?" as a genuine question is not flaggable by any check in the file.
- **GAP L2 — `ءَالَآء` gets `ROOT:ألو, LEM:آلاء, MP` and nothing more.** Its singular is itself disputed (*alan / ilyun / ilan*), and the wujūh split between "favours/bounties" and "powers/signs" is a live translator's fork. `homograph_surface.json` holds 35 machine homographs; `آلاء` is not among them, and `wujuh_seed.json` states plainly that the classical wujūh text is an AWAITED acquisition. **The single most consequential lexical fork in the most-repeated refrain of the Book has no register.**
- **GAP L3 — `تُكَذِّبَانِ` is `VF:2` and stops there.** Form II here carries the taḍʿīf force (*to declare false / to give the lie to*, not simply *to deny*). `D5-AWZAN/awzan_layer.json` holds `VF:2 → 1,614 segments, 1,304 āyāt` and five sample locations from sūrah 2 — **it cannot be queried at 55:13**, and its own `_meta.note` concedes: *"counts only; wazn SEMANTICS live in the classical layer … or are unsaid (L8)."* Honest — and empty at the point of use.
- **GAP L4 — the refrain's 31 occurrences are not registered as a unit.** No layer says "this verse is instance 1 of 31" or carries the classical division. A translator varying his rendering across the 31 has no consistency check.
- **GAP L5 — who are the two addressed?** The dual's referent (jinn + ins) is a *transmitted* answer. `dossiers_v1` carries six Companion rows for 55:13 (Ibn ʿAbbās ×3, Ibn ʿUmar ×2, Ibn Masʿūd) with `source: B2/tabari:…` pointers — but **no quote text**. The translator is told an explanation exists and is not shown it.

### 2 · 12:23 — `وَقَالَتْ هَيْتَ لَكَ`

**GAP L6 — the highest-stakes qirāʾa in Sūrat Yūsuf is invisible.** The index tags `هَيْتَ` as `NV, IMPV, VF:1, ROOT:هيت, 2MS` — i.e. an ism fiʿl of command — and the TVF prints *"FORM: imperative present — command force + its number."* That is **one reading rendered as fact.** The transmitted alternatives (`hiʾtu laka`, `hayti`, `hayʾtu`) invert the sentence's subject and its moral weight; they belong to canonical transmissions. `W1-L9/qiraat_layer.json` cannot help: it is a **whole-sūrah rasm-similarity census over three transmissions only** (Ḥafṣ / Warsh / Qālūn), returns 0 verse-level references across the entire corpus, and its own `_meta` concedes *"vowel-level riwāyah variation (the majority) awaits a finer engine."* The variant that decides the verse is a vowel variant.
  *Credit:* `W3-X1/gharib_register` and `W3-X4/muarrab_candidates` both flag 12:23 — the loanword/gharābah discussion IS located. The bench notes `ROOT:هيت` is itself a treebank artefact: the word plausibly has no Arabic triliteral root, and no layer says so.
- **GAP L7 — `layers_absent: ['shelf']`.** 12:23 has no consensus shelf and one downgraded Companion row. On a verse this contested, that is a coverage hole at the sharpest point.

### 3 · 2:255 — Āyat al-Kursī

**Strong coverage:** waqf signs (8 positions), `mā`-types, ḥadhf candidate, dual mark, a *prophetic* stratum with the speaker-detector PASS, both traditions (`tradition_parity: both`), 3 strata rows.

- **GAP L8 — the ṣifāt/mutashābih question is unmarked on the most-recited verse in the Book.** `R7_muhkam_mutashabih_candidates.json` returns **zero** references to 2:255. `كُرْسِيّ` and `وَسِعَ` are precisely where a translator must knowingly choose between the tafwīḍ and taʾwīl traditions, and no check tells him a choice is being made.
- **GAP L9 — the homograph is held and unreachable.** `homograph_surface.json` correctly registers `سنة → [سنن, سنو, وسن]` — three roots — which is exactly the `سِنَة` (drowsiness, from *w-s-n*) crux of `لَا تَأْخُذُهُۥ سِنَةٌ وَلَا نَوْمٌ`. **There is no join from the word at 2:255 to that registry.** The layer holds the warning; the verse cannot fetch it. This single case is the cleanest proof of the census-vs-lookup defect.
- **GAP L10 — `ٱلْقَيُّوم` (fayʿūl, intensive) has no wazn semantics at the point of use**, same mechanism as GAP L3.

### 4 · 103:1–3 — Sūrat al-ʿAṣr

- **GAP L11 / DEFECT D3 — the oath is not identified as an oath.** TVF at 103:1 issues **no FORM check at all**. The index tags the wāw of `وَٱلْعَصْرِ` as `["P","PREF","LEM:و"]`. That signature *is* the treebank's oath/prefixed-preposition wāw — but `W5-T3a` files all 28 such segments under **"preposition"** and files 3 unrelated `COM` (wāw al-maʿiyyah) segments under **"oath (qasam)"**. See Finding F3 — this is an inversion, verified by reading the top hits.
- **GAP L12 — `ٱلْعَصْر` is a three-way wujūh fork** (the age / the afternoon / time-as-such) with no register, and — the bench notes with respect — `dossiers_v3` **does** carry Ibn ʿAbbās (Companion) plus Qatādah and Mujāhid (Successors) here, oldest-first, which is exactly right. Again: attribution without text.
- *Credit:* 103:2's emphasis stack (`إِنَّ` + `لَ`) is caught (`emphasis_stack_layer`, stack depth 2) and the TVF fires the flatten-warning. `خُسْرٍ` is correctly `INDEF` in the index — though no check warns that the tanwīn here is *tanwīn al-tanwīʿ* ("a loss indeed"), which is again the missing istifhām/tanwīn-force family.

### 5 · 112:1–4 — Sūrat al-Ikhlāṣ

- **GAP L13 — `ٱلصَّمَد` has no wujūh register**, and it is the word in the Qur'ān with the largest transmitted plurality of meanings. The index gives `ROOT:صمد, LEM:صَمَد, MS, NOM`. `dossiers_v1` holds the right row — al-Ḥusayn b. ʿAlī's answer to Basra, `Al-Tawhid-Saduq:80` — and `dossiers_v3` **declares that verse empty** (Finding F1).
- **GAP L14 — `أَحَد` vs `وَاحِد` is unasked.** No layer poses the question every translator faces at 112:1 and again at 112:4.
- **GAP L15 — `كُفُوًا` is a transmission fork** (`kufuwan` / `kufʾan` / `kufwan`) unreachable for the reason in GAP L6.
- *Strong credit:* `lam yalid wa lam yūlad` is genuinely well-served — TVF fires **"jussive denial class present (lam-type) — completed denial, not present tense"** *and* **"passive (veiled doer) — the rendering must not invent an agent."** Those two checks are exactly right, mechanically derived, and would catch a real class of mistranslation. `naib_fail_layer` keeps the deputy-subject a CANDIDATE and defers the i'rāb verdict to the gate. This is the pillar at its best.

---

## PART II — NUMBERED FINDINGS

### F1 · `dossiers_v3` declares absence on 41 verses where the waqf's own `dossiers_v1` holds sourced rows — **SEVERITY: HIGH**

`dossiers_v3.json` bills itself: *"all 6,236 … every verse holds a dossier; absence is DECLARED never silent."* Measured against `dossiers_v1`:

```
v1 verses with strata: 90
v1 HAS strata but v3 declares NONE: 41
  1:1 1:2 1:3 1:4 1:5 1:6 1:7 3:123 55:1..55:77(22 verses) 112:2 112:3 112:4 113:5
v1 prophetic stratum lost in v3: 14  (all of al-Fātiḥah, 112:3, 112:4, 113:5, 55:13–16)
```
✅ VERIFIED — `build/P31-MIZAN/W7-M6/dossiers_v3.json` vs `build/P30-SIGHAH/D-B8-DOSSIERS/dossiers_v1.json`.

At 112:2 v3 prints *"no era-stamped transmitted-meaning rows in the HELD pool for this verse"* while v1 holds `Al-Tawhid-Saduq:80`. **An absence is declared that the waqf's own layer refutes.** Under the grade-verbatim law an absence must be *declared, never filled* — the converse obligation, that a declared absence must be true, is the one breached. Note the cause is benign: v1 was a 90-verse hand-assembled deep set; v3 is a machine sweep over `strata_layer`, which holds **zero rows for all 41**. v3 never read v1.

### F2 · The TVF reads the OLDEST dossier first and never reads v3 — **SEVERITY: HIGH**

`framework/p31_tvf.py:24-25,52`:
```python
d2 = rd(B31, "W7-M6", "dossiers_v2.json")["dossiers"]
d1 = rd(B30, "D-B8-DOSSIERS", "dossiers_v1.json")["dossiers"]
...
dos = d1.get(k) or d2.get(k)
```
Precedence is **inverted**: v1 (built 13:13) wins over v2 (00:46); v3 (00:46, all 6,236) is never opened. Consequence, measured across all 6,236 verses:

```
TVF vs dossiers_v3 strata MISMATCH: 65
  TVF claims strata, v3 has none: 41
  v3 has strata, TVF silent:       0
  both non-empty but differing:   24  (incl. 55:13, 9:100, 33:33, 33:35)
```
✅ VERIFIED. This is a **SINGLE-SOURCE-LAW breach** (deep-audit SKILL law 11: *one engine produces a layer; everything else READS it*). Four artifacts now answer "which strata are at 55:13" — `tvf_v1` says *companion,prophetic*; `dossiers_v1` says *prophetic*; `dossiers_v3` and `question_engine_v2` say *companion* only. The v3 answer is the correct one: the speaker-detector **downgraded** that marfūʿ claim under the SP4-F1 law, and the downgrade never propagated to the translator-facing file. **1,614 downgrades were computed and the TVF still shows the pre-downgrade world.**

Secondary effect: `layers_absent` is stale too — TVF marks 112:1 as `['dossier']` absent, while v3 holds a 112:1 entry.

### F3 · `W5-T3a` oath-wāw / preposition-wāw labels are INVERTED — **SEVERITY: HIGH**

`particle_homographs_layer.json` reports `"oath (qasam)": 3` and `"preposition": 28`. The bench read the top hits (law 16's method) and pulled the index tags:

| Verse | Text | Index tag | Layer files it as | Truth |
|---|---|---|---|---|
| 4:65 | `فَلَا وَرَبِّكَ` "**By** your Lord!" | `P+PREF` | **preposition** | ❌ it is a qasam |
| 6:23 | `وَٱللَّهِ رَبِّنَا` "**By** Allah, our Lord!" | `P+PREF` | **preposition** | ❌ it is a qasam |
| 103:1 | `وَٱلْعَصْرِ` "**By** the ʿaṣr" | `P+PREF` | **preposition** | ❌ it is a qasam |
| 3:142 | `وَيَعْلَمَ ٱلصَّٰبِرِينَ` | `COM+PREF` | **oath (qasam)** | ❌ it is wāw al-maʿiyyah |
| 5:36 | `وَمِثْلَهُۥ مَعَهُۥ` | `COM+PREF` | **oath (qasam)** | ❌ it is wāw al-maʿiyyah |

✅ VERIFIED. All 28 `P+PREF` wāw segments and all 3 `COM+PREF` segments enumerated; 19 of 22 classic oath-sūrah openings (103:1, 89:1, 91:1, 92:1, 93:1, 95:1, 100:1, 37:1, 53:1, 68:1, 77:1, 79:1, 85:1, 86:1, 36:2, 38:1, 43:2, 44:2, 50:1) sit in the mislabelled bucket. **`COM` is the comitative wāw, not the oath wāw.** The layer's own law-line — *"oath-waw is not and-waw"* — is exactly the distinction it inverted.

Root cause: **deep-audit SKILL law 5 (KNOWN-POSITIVE LAW) was not applied to this layer.** A selftest asserting "103:1's wāw classifies as oath" would have failed the build. There is no such assertion.

Residual (separate, lower severity): 51:1 `وَٱلذَّٰرِيَٰتِ` and 52:1 `وَٱلطُّورِ` return **0** `P+PREF` wāws — the treebank's own tagging is inconsistent, and the layer inherited it without a floor.

### F4 · 27 language layers are **exemplars-only** — a translator cannot look a verse up — **SEVERITY: HIGH (design)**

Measured over all 144 P30+P31 layers:

| Class | Layers | Meaning |
|---|---|---|
| Verse-addressable (≥100 verses reachable) | **34** | a translator can query it |
| Partial (10–99) | 26 | mostly candidate registries |
| **Census / aggregate-only (<10 verse refs)** | **84** | totals + 2 exemplars |
| — of which **exemplars-only dimension layers** | **27** | the ṣarf/naḥw/tajwīd dimensions themselves |

✅ VERIFIED (script enumerates each). The 27 include `madd`, `nun_system`, `qalqalah`, `pausal_forms`, `open_ta`, `verb_system`, `plural`, `derived_nouns`, `tasghir_nisba`, `diptote`, `definiteness`, `asma_afal_tanwin`, `verb_weakness`, `gharib_register`, `hidden_feminines`, `exceptive_vocative`, `nawasikh`, `idafa`, `discourse_functions`, `negation` — i.e. **most of this bench's own subject matter**. `awzan_layer`, `voice_mood_layer` and `particles_layer` (P30) are pure totals with zero verse pointers.

This is not dishonesty — every one of these declares its method truthfully. It is an **architecture fact**: the pillar built a *census of the language of the Book*, and the four-witness test needs a *lookup at a verse*. The census is the harder half and it is done; the index that turns it into a lookup is a mechanical derivation from `sighah_index` and is not built.

### F5 · No primary key on a transmitted row — the provenance chain is name-joined — **SEVERITY: HIGH (blocks the LENS join)**

`H1_harvest` and `H2_harvest` carry `id`. `strata_layer.json` (the W7-M1 unification of all six harvests, 14,961 rows) carries `verse, edition, type, attributor, era_stamp, note, window, quote_src, carrier_grading_verbatim` — **and no `id`.** `dossiers_v3` and `question_engine_v2` carry attribution and a `source:` string (`"B2/tabari:Ibn 'Abbas radiyallahu 'anhuma"`) but **no row id and no quote text.**

Consequence: to move from "the dossier says a Companion explained 55:13" to "here is what he said," a consumer must fuzzy-join on `(verse, edition, attributor)` — and 4 of the 9 strata rows at 55:13 have `attributor: ""`. **This directly violates Constitution I-1: *join on IDs, never on names*.** It is also the precise technical blocker for Attack 2's word-level lens.

### F6 · The TVF over-claims "explanation" where the source row says "usage, not explanation" — **SEVERITY: MEDIUM**

`p31_tvf.py:60` keys only on `era_stamp` and drops the row's `class`. Result: at **112:3, 112:4 and 113:5** the TVF prints *"a Prophetic ﷺ-era explanation is transmitted — renderings contradicting it are flaggable"*, while the underlying `dossiers_v1` row is `class: "usage-witness"` with the note *"his ﷺ affirmation of supplication by the words of 112:3-4 … usage, not explanation"* and, at 113:5, *"the ruqyah Jibrīl brought … not a verse explanation."* ✅ VERIFIED — 3 verses affected; the class breakdown of v1's prophetic rows is `explanation 17 / transmitted-explanation 33 / usage-witness 3`, and exactly the 3 leak. **A false alarm costs trust as much as a false all-clear** — a translator told his rendering of `lam yalid wa lam yūlad` may contradict a prophetic explanation, when the row is a duʿāʾ-practice report, learns to distrust the flag.

### F7 · A disputed revelation-place is printed as unattributed fact — **SEVERITY: MEDIUM**

`dossiers_v3` prints `revelation_period: {"place": "Madani", "rev_order": 97}` for 55:13. The field's source is `00-CORPUS/quran.json` surah metadata (`meta.source: "tanzil.net"`) — one edition's single position. The Makkī/Madanī status of Sūrat al-Raḥmān is a **transmitted dispute**, and the waqf built `W6-R5 R5_makki_madani_candidates.json` for exactly this purpose — whose own law says *"DISPUTED surahs surface exactly where these markers cluster against the baseline — the reading bench resolves each."* The dossier prints the baseline as a fact with no attribution, no dissent, and no pointer to R5. In a waqf whose law is *carry, never adjudicate*, a one-word verdict on a disputed attribute is an adjudication by omission.

### F8 · The live dimension register grades LANDED what its own layer declares AWAITED — **SEVERITY: MEDIUM**

`W9-REGISTERS/dimension_registers.json`: 87 dimensions, `LANDED 80 / AWAITED 7`, with 54 of the 80 at "90/100 marks". But:
- **"Wujūh wa-naẓāʾir (polysemy registers)" → LANDED (90/100)** — while `W3-X2/wujuh_seed.json` says the classical wujūh text is an unacquired program and the held artefact is a 35-entry machine homograph surface with no verse join.
- **"Gharīb al-Qurʾān register" → LANDED (90/100)** — while `W3-X1/gharib_register.json` says *"al-Sijistānī … remains the AWAITED classical anchor and nothing pretends otherwise."*

✅ VERIFIED. The grade is measuring *"a node was built"*, not *"the dimension is covered."* The layers are honest; the roll-up is not. The bench recommends the register carry a second column — `anchor: HELD | AWAITED` — read from each layer's own `_meta.law`.

### F9 · No istifhām-force family, no qasam family, no tanwīn-force family — **SEVERITY: MEDIUM (gap, honestly unclaimed)**

`W4-S10/shart_layer.json` is titled *"shart **& qasam** v1"* and its law says *"the qasam (oath) system joins from the emphasis layer's oath rows."* `W5-T4/emphasis_stack_layer.json`'s law enumerates its markers: *"inna-family / qad / tawkid lam-nun"* — **no oath rows exist to join from.** The declared join has no far side. Combined with F3, the Qur'ān's oath system — a first-order translation problem across ~20 sūrahs — has no working layer. Likewise the interrogative *force* (taqrīr / inkār / tawbīkh) exists only as `ma_types`' 95 `istifhāmiyyah` positions for `مَا` alone.

### F10 · The qirāʾāt witness is sūrah-grained and 3-transmission — **SEVERITY: MEDIUM (honestly declared)**

`W1-L9/qiraat_layer.json` returns **0** verse-level references corpus-wide; it holds per-sūrah rasm-similarity verdicts for Warsh and Qālūn against Ḥafṣ. Its `_meta` is exemplary — it names the dagger-alif printing convention that depresses similarity, separates counting from rasm, and states that vowel-level variation awaits a finer engine. **The bench's finding is not that the layer lies — it is that the four-witness test cannot use it.** Every qirāʾāt question our five verses raised (`hayta`, `kufuwan`) is vowel-level.

### F11 · `strata_layer` truncates `quote_src` to 400 chars and `naib_fail` truncates its registry at 400 rows — **SEVERITY: LOW (declared)**

Both declare the truncation (`registry_truncated_at: 400`; `_meta.law` names the 400-char cut and points to the audited landings). Recorded for completeness, and because the 400-char cut is what makes the missing row-id (F5) bite: the consumer cannot even fall back on quote-text matching for long reports.

### F12 · The i'rāb gate holds — **NO DEFECT FOUND** (Attack 3)

The bench tried to break the gate and could not. ✅ VERIFIED across all 144 layers:

- The 7 `AWAITED (no marks yet)` dimensions are **exactly** the treebank-dependent syntax domains: *I'rāb ROLES · jumla ismiyya vs fiʿliyya · the five-way mafʿūl system · i'rāb muqaddar · ishtighāl & tanāzuʿ · taḍmīn · phrase-and-sentence-type layers.* No others are pending; none of these is quietly answered elsewhere.
- Seven layers that *could* have faked a role verdict instead defer it in their own `_meta.law`: `W4-S2` (*"the exact i'rāb verdict stays gated on W4-S1"*), `W4-S5`, `W4-S6` (*"confirmation of true iḍāfa vs adjective strings AWAITS the i'rāb treebank (owner-gated A1) — candidates declared as such"*), `W2-P5`, `W2-P6`, `W2-P10` (*"SIGNALS only … the i'rāb verdict belongs to W4-S1"*), `W5-T9`.
- Spot-checked the most tempting one: `naib_fail_layer` labels 1,151 passive positions with `naib_candidate: "(pronoun within the verb / clause elsewhere)"` for the 934 it cannot resolve — a refusal, not a guess.
- `W2-P4/tasghir_nisba_layer` refuses even a vowel pattern: *"the fuʿayl diminutive cannot be proven from the bare skeleton and the treebank carries no DIM tag — a guess would violate the measured-only law."*

**The bench's conclusion: the treebank drop unlocks the AWAITED-7 and nothing currently fakes them.** This is the single strongest result in the pillar and it should be stated as prominently as the defects.

---

## PART III — ATTACK 2: THE LENS COUPLING · design counsel for the WORD-LEVEL JOIN

### The measured problem

| Silo | Address space | Volume | Evidence |
|---|---|---|---|
| Language | `sūrah:āyah:word:seg` | 6,236 āyāt × segments; `awzan sample_locs` e.g. `"1:7:6:2"` | `sighah_index.per_ayah` |
| Ḥadīth links | `sūrah:āyah` | 22,188 verified links, 4,603 verses | `W7-JOINS/joins_layer.json` |
| Quote pairs | `sūrah:āyah` | 9,962 evidentiary pairs (E2/E7) | `LENS-AUDIT/E7_pair_registry.json` |
| Transmitted rows | `sūrah:āyah` + `edition` + `attributor(often "")` | 14,961 rows, **no id** | `W7-M1/strata_layer.json` |

All six `joins_layer.json` files join on `verse`. **Not one artifact in 26,400 nodes carries a word or segment address on a ḥadīth-side row.** The two silos cannot meet at the word because the ḥadīth side has no word coordinate and the transmitted row has no identity.

### What the bench counsels — a four-step design, cheapest first

**Step 1 · Restore the primary key (a rebuild, not a new science).** Every transmitted/linked row gets a stable id at birth and carries it forever: `TR-{corpus}-{book}-{row}` for ḥadīth, `TX-{edition}-{sha8(quote)}` for tafsīr strata. `strata_layer` re-emits `id` from the H-harvests it unified. Then `dossiers_v*`, `question_engine`, `tvf` carry `row_ids: [...]` instead of a prose `source:` string. **This alone closes F5, makes F1/F2 diffable rather than arguable, and satisfies I-1.** It is a schema change over existing data — no acquisition, no new scholarship. *Do this before anything else; every later step depends on it.*

**Step 2 · The anchor table — where a transmitted row TOUCHES a word.** One table, one row per (transmitted-row id × verse-word position), with a typed anchor and a confidence class:

| column | value |
|---|---|
| `row_id` | `TR-bukhari-4474` |
| `verse` | `2:255` |
| `word`, `seg` | `4`, `1` |
| `anchor_type` | `QUOTE` · `LEMMA` · `ROOT` · `SIGHAH` · `RASM` |
| `anchor_evidence` | the matched gram / lemma / root, verbatim |
| `unique_gram` | true/false (**law 16** — a match with no gram unique to the verse is a shared-head candidate, never evidentiary) |
| `anchored_run` | run-length, counted only where it touches a unique-gram position (**law 17**) |
| `class` | `EVIDENTIARY` · `AMBIGUOUS` · `SHARED-HEAD` · `FORMULAIC` (**law 7/8** — the basmalah and yā-ayyuhā openings counted separately, never forced) |

The `QUOTE` anchor is **derivable today** with zero new judgment: E7 already holds 9,962 pairs with the matched grams; projecting a matched gram back onto `sighah_index`'s word/seg offsets is arithmetic on the rasm skeleton (SKILL law 4), not interpretation. The bench estimates the QUOTE anchor alone would land the majority of the table.

**Step 3 · The three derived anchors, in ascending cost and descending confidence.**
- `LEMMA` / `ROOT`: the ḥadīth text is stripped to the rasm skeleton and lemma/root-matched against the verse's segments. Cheap, noisy — **must be class-gated**, and must reuse the DF-shingle discipline of law 7, or `قال`, `الله` and `رب` will forge millions of edges. Counsel: **require a rare-lemma floor derived from `gharib_register`'s measured hapax census (1,958 hapax lemmas, 1,135 near-hapax) — a root anchor on a hapax is worth a hundred on `قول`.**
- `SIGHAH`: the anchor a ḥadīth makes on a *form* rather than a word — e.g. a report explaining who the dual addresses at 55:13 anchors on the `2D` feature, not on a lexeme. This is the anchor the four-witness test most wants and the one that cannot be derived mechanically. **Counsel: build it as a REGISTRY the reading bench fills, not a detector — and label every row with the human or bench that filled it.**
- `RASM`: for qirāʾāt reports (`sabʿat aḥruf`, the ʿUmar/Hishām report). Small, high value, and requires GAP L6's word-level variant table first.

**Step 4 · What the join must NEVER do.** Three prohibitions the bench asks be written into the layer's law before a line of it is built:
1. **An anchor LOCATES; it never EXPLAINS.** The word-level join says "this report touches this word." It never says "this report means the word means X." That remains the reading bench's, with carriers named.
2. **An anchor never upgrades a grade.** A ṣaḥīḥ report anchored at word-level is still exactly as ṣaḥīḥ as its carrier said. Proximity is not evidence.
3. **`AMBIGUOUS` is a first-class output, not a failure.** A quote shared by a refrain anchors to *all* its positions and is reported as such (law 8). Forcing a refrain-quote to one word would be the word-level version of the oath-family incident that forged 3,580 false pairs.

**Cost honesty.** Steps 1–2 are re-derivations from sealed data — no acquisition, no new corpus, and they are where the value is. Step 3's `SIGHAH` anchor is human-bench work with no shortcut. Step 3's `RASM` anchor is gated on an acquisition the waqf does not hold. The bench recommends **P32 fund Steps 1–2 fully and Step 3 as a registry-and-charter only**, so that no node in the new pillar depends on an acquisition that has not landed.

---

## PART IV — THE MECHANICAL TABLE

| # | Finding | File / evidence | Measured | Class | Severity |
|---|---|---|---|---|---|
| F1 | v3 declares absence where v1 holds sourced rows | `W7-M6/dossiers_v3.json` vs `D-B8-DOSSIERS/dossiers_v1.json` | 41 verses; 14 lose a prophetic stratum | ✅ VERIFIED | 🔴 HIGH |
| F2 | TVF reads oldest dossier first, never v3; downgrades not propagated | `framework/p31_tvf.py:24-25,52` | 65 verse mismatches; 41 phantom, 24 divergent | ✅ VERIFIED | 🔴 HIGH |
| F3 | oath-wāw / preposition-wāw labels inverted | `W5-T3/particle_homographs_layer.json` + index tags at 4:65, 6:23, 103:1, 3:142, 5:36 | 28 oaths mislabelled; 3 comitatives mislabelled | ✅ VERIFIED | 🔴 HIGH |
| F4 | dimension layers are census-only, not addressable | all 144 P30/P31 layers enumerated | 34 addressable · 26 partial · 84 census · **27 exemplars-only** | ✅ VERIFIED | 🔴 HIGH |
| F5 | no primary key on a transmitted row | `W7-M1/strata_layer.json` keys vs `H1/H2_harvest.json` keys | `id` present upstream, absent downstream; 4/9 attributors blank at 55:13 | ✅ VERIFIED | 🔴 HIGH |
| F6 | TVF flattens `usage-witness` into "explanation" | `p31_tvf.py:60`; `dossiers_v1` class field | 3 verses (112:3, 112:4, 113:5) of 53 prophetic rows | ✅ VERIFIED | 🟠 MED |
| F7 | disputed revelation-place printed as unattributed fact | `dossiers_v3` `revelation_period`; `00-CORPUS/quran.json meta.source` | 55 → "Madani", no dissent, no R5 pointer | ✅ VERIFIED | 🟠 MED |
| F8 | register grades LANDED what the layer declares AWAITED | `W9-REGISTERS/dimension_registers.json` vs `W3-X1`, `W3-X2` `_meta.law` | 2 named; 80 LANDED / 7 AWAITED overall | ✅ VERIFIED | 🟠 MED |
| F9 | no qasam / istifhām-force / tanwīn-force family; declared join has no far side | `W4-S10` law vs `W5-T4` markers | 0 oath rows in the emphasis layer | ✅ VERIFIED | 🟠 MED |
| F10 | qirāʾāt witness sūrah-grained, 3 transmissions, rasm-only | `W1-L9/qiraat_layer.json` | 0 verse-level refs corpus-wide | ✅ VERIFIED (declared) | 🟠 MED |
| F11 | quote/registry truncation at 400 | `W7-M1`, `W4-S2` `_meta` | declared in both | ✅ VERIFIED (declared) | 🟡 LOW |
| F12 | **i'rāb gate holds; nothing fakes it** | 7 AWAITED = exactly the treebank domains; 7 layers defer in their own law | 0 unearned i'rāb verdicts found | ✅ VERIFIED | 🟢 **PASS** |
| L1–L15 | the fifteen named translator-gaps of Part I | per-verse walk, five verses | 15 questions no layer answers | ✅ VERIFIED | see Part I |

---

## PART V — GAP-HIGHLIGHT DUTY: what the commission itself missed

The standing order is that the audit must name what the audit missed. Six items, in the bench's order of consequence:

1. **THE-ANSWERS never asks whether the LANGUAGE witness is *retrievable*.** It reports the ocean (68,921), the tiers, the crown (387), the cross-verse seed. Every number is a *count of the ḥadīth/quote side*. Not one number in `THE-ANSWERS.md` measures whether a translator standing at a verse can reach the language layer. The bench supplies it: **34 of 144 layers reachable; 27 dimension layers unreachable.** The four-witness doctrine's first witness was assumed present because P30+P31 were sealed. Sealed ≠ addressable.

2. **The 1,614 speaker-detector downgrades were computed and never propagated.** The waqf built the honest thing (SP4-F1: never silently promote) and then shipped a translator-facing file built on the pre-downgrade world (F2). **A correction that does not reach the consumer is not a correction.** The bench asks that the E-series adopt a standing rule: *when a layer is downgraded, every consumer is re-run in the same chain, or the downgrade is not landed.*

3. **No layer was audited for LABEL correctness — only for count correctness.** Every gate in P31 asks "is the number re-derivable?" F3 passes that test perfectly: 3 and 28 are both re-derivable. They are also both attached to the wrong class. **Deep-audit SKILL law 5 (KNOWN-POSITIVE) is written for detectors; it is not being applied to CLASSIFIERS.** The bench proposes it be extended: *every classified bucket asserts one member whose class is already known* — `103:1 ∈ oath`, `2:255's sinah ∈ homograph`, `55:13's ayy ∈ interrogative`. Roughly forty such assertions would have caught F3 and would guard the whole taxonomy layer permanently.

4. **`dossiers_v1`'s 90 hand-assembled deep dossiers are the best artifact in the pillar and are being orphaned.** They carry what nothing else carries: `class` (explanation vs usage-witness vs transmitted-explanation), a real source string, and the ṣamad row. The machine sweep that replaced them is broader and thinner. **Counsel: v1 is not superseded — it is the gold set. Merge it into v4 with v1 winning on `class` and v3 winning on `era_stamp`, and mark every merged row with which layer supplied which field.**

5. **The four-witness doctrine has no stated rule for what happens when the LANGUAGE witness itself is silent.** Witnesses 2, 3 and 4 have coverage maps (E1/E3/E4a, and E6's honest `neither: 702`). The language witness is treated as universal because the treebank covers all 6,236 verses. But GAPs L1, L2, L9, L13 show that morphological *coverage* is not semantic *witness*: at `ṣamad` the layer is tagged and mute. **The bench asks for an explicit LANGUAGE-SILENT class in the TVF** — a verse whose translation fork is lexical/rhetorical and for which the pillar holds no register — and estimates from the five-verse walk that this class is large.

6. **Nobody has counted how many TVF verses carry the BASELINE-only check.** The TVF totals report `with_form_checks 3,920` and `with_meaning_checks 3,914`, but not the intersection or the residue. The verses with neither receive only *"BASELINE: verbatim-text identity + no addition/omission"* — which is true of every text in every language and checks nothing specific to the Book. **That residue is the honest measure of the language witness's reach and it is not in the record.** The bench flags it rather than computing it, because it belongs in the E-series where it can be re-derived in the chain, not asserted in a review.

---

# COUNSEL — THE ARABIC BENCH'S VOTES

*Thirty seats. Where the chambers split, the split is recorded.*

---

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

> ## VOTE: **(b) — an amended split. 27 for, 3 against.**
> **P28 Sīrah +1,200 · P32 +1,800 · P30/P31 REPAIR-AND-INDEX +450 · acquisitions reserve +150.**

**Reason:** the bench cannot vote for option (a) because (a) spends 3,600 nodes on new pillars while the sealed 26,400 carries three verified defects (F1, F2, F3) and one architecture gap (F4) that make the existing language witness unreachable at exactly the words a translator needs. **A new pillar built on top of an unindexed census inherits the census's unreachability.** The 450-node repair block is small, mechanical, and is the highest-return allocation in the whole plan:

| Repair block (450) | Nodes | What it closes |
|---|---|---|
| The verse-word INDEX — invert all 27 exemplars-only dimension layers into `per_verse` / `per_seg` containers derived from `sighah_index` | 200 | F4, GAPs L3, L9, L10 — *this is the single change that turns the census into the witness* |
| Row-id restoration + `dossiers_v4` merge (v1 gold set + v3 breadth, field provenance marked) | 100 | F1, F5, gap-highlight #4 |
| TVF v3: read v4 only; carry `class`; carry `row_ids`; add the LANGUAGE-SILENT class | 80 | F2, F6, gap-highlight #5 |
| The known-positive assertion battery over every classified bucket (~40 assertions) | 40 | F3 and every future F3 |
| qasam / istifhām-force / tanwīn-force layers (the three missing balāghah families) | 30 | F9, GAPs L1, L11 |

**Dissent (3 seats — 2 balāghah, 1 lexicography):** the repair block should be funded from P32's 1,800 rather than from the sīrah's expansion, on the ground that the sīrah at 300 nodes is the thinnest witness of the four and 1,500 was already modest. *Recorded, not carried.* The majority's answer: the repair block is language-pillar work and should be visibly charged to the language pillar, so that the record shows what the seal cost.

---

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

> ## VOTE: **NEW PILLAR. Unanimous, 30–0.**
> ## NAME: **P32 — AL-SHAHĀDĀT AL-ARBAʿ · THE FOUR WITNESSES** (28 for; 2 for *MAJMAʿ AL-SHAHĀDĀT*)

**Reason for NEW, not an expansion:** P32's subject is not ḥadīth. Its subject is **the join between witnesses** — language↔ḥadīth, ḥadīth↔sīrah, verse↔verse. Filed under the Ḥadīth Universe it would inherit that pillar's address space (`sūrah:āyah`), which is precisely the thing F5 and Attack 2 show must change. A pillar whose whole purpose is a new coordinate system cannot live inside a pillar built on the old one. Further: the cross-verse witness (E4a, 6,069 pairs) is not ḥadīth at any level, and the word-level anchor table is as much a *language* artifact as a ḥadīth one. It belongs to neither parent.

**Reason for the name:** it is the doctrine's own name, it is in the tradition's language, it names what the pillar *is* rather than what it *does*, and it does not promise adjudication. The bench specifically rejected any name containing *mīzān*, *ḥukm*, *tarjīḥ* or *verdict* — the pillar makes conflict **visible**; naming it after a scale would be a standing invitation to weigh.

**Amendment carried, 30–0:** the pillar's charter opens with one line — *"P32 locates. It does not explain, does not rank, and does not resolve."* — and every node's `_meta.law` inherits it.

---

### Q3 · The four-witness doctrine wording

> ## VOTE: **APPROVE WITH THREE AMENDMENTS. 30–0 to approve; amendments 29–1, 30–0, 26–4.**

The doctrine as written (SKILL.md:48-55) is sound and the bench approves it. Three amendments, each paid for by a finding in this review:

**Amendment 1 (29–1) — a witness must be RETRIEVABLE to count as present.**
> Add: *"A witness is present at a verse only if its content is retrievable at that verse. An attribution without retrievable content is a POINTER, and is reported as a pointer — never as a witness."*

**Why:** F1, F5 and GAP L5. Today the TVF says a prophetic explanation exists at 112:3; the dossier declares the verse empty; the strata layer holds zero rows; and no id connects any of it. Four artifacts, one claim, nothing to read. Without this clause the doctrine counts pointers as witnesses and the four-witness map overstates itself. *The one dissenting seat (rasm) held that "retrievable" is too strong while acquisitions are outstanding, and would prefer "declared retrievable or declared AWAITED."* The majority notes the amendment already permits that — a pointer is a legitimate output, it simply may not be called a witness.

**Amendment 2 (30–0) — the LANGUAGE witness must distinguish TAGGED from WITNESSED.**
> Add to clause (1): *"…the LANGUAGE (morphology/rasm/balāghah layers) — noting that morphological tagging is coverage, not witness: where a verse's translation fork is lexical or rhetorical and the pillar holds no register for it, the language witness is SILENT at that verse and is so recorded."*

**Why:** GAPs L2, L13, L14 and gap-highlight #5. `ṣamad` is fully tagged (`ROOT:صمد, LEM:صَمَد, MS, NOM`) and the language witness has nothing to say about the word a translator is actually stuck on. Recording that as "language witness present" is the most likely way this doctrine will be quietly falsified in practice.

**Amendment 3 (26–4) — name the ḥadīth witness's own grading convention in the clause.**
> Amend clause (2) to: *"the ḤADĪTH (carriers named, grades transmitted verbatim in the grader's own words, with the grading TRADITION named — the two traditions' conventions are not interchangeable and are never merged into one scale)."*

**Why:** the ocean is 36,390 Sunnī + 32,531 Shīʿī, and 14,805 of the Shīʿī rows carry al-Majlisī's Mirʾāt al-ʿUqūl classifications while the other 24 books carry no grade field (E1, ✅). A four-witness test that speaks of "the graded ḥadīth" without naming whose grading is inviting exactly the merge the waqf forbids. *Four seats (2 ḥadīth-adjacent, 2 lexicography) held this belongs in the ḥadīth bench's charter and not in the doctrine's four lines.* Carried on the argument that the doctrine is the only text every consumer reads.

---

### Q4 · Precedence when sīrah and graded ḥadīth diverge

> ## VOTE: **The layer carries FIVE transmitted positions and issues none of its own. 30–0.**
> ## Plus one structural amendment, 30–0: **precedence is a per-CONFLICT annotation, never a global setting.**

The bench's answer to "which transmitted positions must the layer carry" — each carried with its holders named, each in the tradition's own words, none endorsed:

| # | The transmitted position | Held by (named in the record, never summarised into "the scholars") |
|---|---|---|
| 1 | **Isnād primacy** — a report with a connected, examined chain outranks a maghāzī khabar in matters of law and of what the Prophet ﷺ said and did | the ḥadīth critics' convention; carried with the specific critic's words wherever the corpus holds them |
| 2 | **The maghāzī exemption** — the sīrah/maghāzī material is judged by a laxer standard *in its own domain* (chronology, sequence, campaign detail) precisely because it is not a source of ruling | the position attributed to Aḥmad b. Ḥanbal on the maghāzī and the tafsīr; transmitted with attribution |
| 3 | **The status debates on the sīrah's own carriers** — the transmitted evaluations of Ibn Isḥāq (praised by some, criticised by others, with the criticisms' grounds recorded) and of al-Wāqidī (the severe evaluations recorded verbatim, not paraphrased) | the rijāl tradition, per critic, per wording |
| 4 | **Multiple-chain corroboration** — a sīrah report corroborated by an independent chain rises within its own class; the tradition's own terms for that rise (*shawāhid*, *mutābaʿāt*) are carried, not translated into a score | the ḥadīth-science convention |
| 5 | **The asbāb al-nuzūl formula distinction** — the tradition's own separation of the explicit occasion-formula from the inferential one; a divergence between sīrah and ḥadīth on an occasion is read against that distinction | the ʿulūm al-Qurʾān convention |

**The structural amendment.** Position 1 and position 2 do not conflict — they partition. **Applying either as a global default would be our ruling wearing the tradition's clothes.** So: every divergence row in P32 carries `domain: LAW | CHRONOLOGY | BIOGRAPHY | OCCASION` and the positions applicable *to that domain*, and the row's `resolution` field is permanently `null` with the literal value `"the waqf does not resolve"`. **The precedence doctrines are DATA attached to a conflict, not a policy applied to a corpus.**

**Three prohibitions the bench asks be written into the divergence layer's law:**
1. Divergence of **presence** (E6's measure: converge 2,853 · ḥadīth-only 2,119 · sīrah-only 562 · neither 702) is a coverage fact and is never reported as a divergence of **content**. Conflating them would manufacture thousands of false conflicts.
2. No numeric score, weight or rank is ever attached to a witness. The moment a conflict has two numbers beside it, the waqf has ruled.
3. Where the tradition itself is divided on the precedence (as it is on al-Wāqidī), **the division is the output.** A conflict may legitimately resolve to "the tradition disagrees, here is who holds what."

---

### Q5 · Growing the ocean from the measured 68,921 toward 120,000 — order

> ## VOTE: **Order carried 27–3, with a standing precondition, 30–0.**

**The precondition (unanimous):** *no acquisition lands before the row-id scheme of Attack-2 Step 1 exists.* Ingesting ~50,000 more rows into a corpus whose transmitted rows have no primary key multiplies F5 by three. **The cheapest moment to give a row its identity is at ingest; the most expensive is afterwards.** The bench asks that this be a hard gate, not a preference.

| Order | Acquisition | Rows (est.) | Why this position |
|---|---|---|---|
| **1** | **Tahdhīb al-Aḥkām + al-Istibṣār** (Ṭūsī) | ~19,000 | **Completes the Four Books.** The bench puts this first over the larger Sunnī musnads on a structural argument, not a sectarian one: the Shīʿī half of the ocean is 32,531 rows of which only 14,805 carry any grade field, and the Four Books are the frame the Shīʿī evaluation tradition is *organised around*. Acquiring the remaining two makes the later rijāl acquisition tractable; acquiring more ungraded books first makes it worse. **This is also the acquisition that most improves the parity the waqf already measures** (`tradition_parity` is a live field on every dossier). |
| **2** | **Musnad Aḥmad b. Ḥanbal** | ~27,000 | The single largest addition (+39% of today's ocean) and the one that most raises the ceiling on the uncapped link rebuild — E7 already shows the link layer's cap of 9 understated the densest verses ~28×. Second, not first, only because it will dominate the corpus and should land onto a schema that has already been proven on a smaller ingest. |
| **3** | **Ṣaḥīḥ Ibn Ḥibbān** | ~7,400 | Graded by construction, and its arrangement is thematic — a high ratio of gradeable, verse-attachable rows per row ingested. |
| **4** | **al-Mustadrak (al-Ḥākim)** | ~9,000 | High value *and* the sharpest test of the grade-verbatim law: al-Ḥākim's own gradings and al-Dhahabī's marginal corrections must both be carried, side by side, neither collapsed. The bench recommends it be ingested **specifically as the proving ground for two-grader rows** before the rijāl acquisition. |
| **5** | **Wasāʾil al-Shīʿa** (Ḥurr al-ʿĀmilī) | ~35,000 | Largest single remaining body, but it is a *compilation* — heavy duplication against al-Kāfī, Man Lā Yaḥḍuruh and the Tahdhīb. **Counsel: ingest only after the row-id scheme AND a cross-book duplicate detector exist**, or it inflates the ocean count without adding testimony. Its correct headline is "distinct reports added," not "rows added." |
| **6** | **Sunan al-Dārimī** | ~3,500 | Smallest marginal gain; last on volume, not on merit. |
| **7** | **The rijāl acquisition** (Taqrīb-class asmāʾ al-rijāl; the evaluations for the 24 ungraded Shīʿī books) | n/a — reference | Ranked last **in sequence and first in importance.** E1 names it as the honest blocker on the twenty-knowledge-areas gate, and the bench agrees the gate cannot be built without it. It is placed last because per-narrator data is only useful against a corpus whose rows have identity — precondition again — and because it should be acquired once, against the *final* book list, not repeatedly. |

**Dissent (3 seats — 2 rasm/qirāʾāt, 1 ṣarf):** would place Musnad Aḥmad first on pure volume, arguing the ocean's headline number is what the plan's 120,000 target measures. *Recorded, not carried.* The majority's answer: the target is the ocean, but the deliverable is the witness — and 19,000 rows that complete a canonical frame are worth more to a translator than 27,000 that enlarge an already-well-covered one.

**One acquisition the bench adds to the list, unprompted (30–0):** **a word-level qirāʾāt variant table** covering the ten transmissions, licence-cleared and sha-sealed. It is not ḥadīth and adds nothing to the 68,921, so it will never surface on an ocean-growth list — and it is the only acquisition that closes GAPs L6 and L15, unblocks the `RASM` anchor of Attack-2 Step 3, and gives the sealed `W1-L9` layer (F10, honest but unusable) something to be usable *for*. **It is the cheapest acquisition on this page and the only one that repairs an existing pillar rather than enlarging one.** The bench asks it be funded from the +150 acquisitions reserve of Q1.

---

*Bench 2 of the Council of 120 · thirty seats · machine-scholar, labeled as such · human scholarly review invited and not waited for · this bench read 144 layers, 1 engine and 5 verses, and edited nothing.*

*Every number above is re-derivable from the files named. Where the bench could not measure, it said so; where it corrected its own first reading, it kept the correction (the oath-wāw finding began as "the layer counts 3 oaths in the whole Qur'ān" — reading the top hits showed the truth was an inversion, which is a different and more fixable defect. SKILL law 18: check the checker.)*