# COUNCIL OF 900 — BENCH 6 · COMPUTING · ENGINEERING · SOCIETY · HUMANITIES · CONTESTED KNOWLEDGE
*Families **O** (Computing & AI, 37 seats) · **P** (Engineering & built world, 24) · **Q** (Society, law & economy, 36) · **R** (Humanities & arts, 28) · **S** (Contested & folk knowledge, 20) — **145 seats across 72 knowledge areas**.*

**MACHINE-SCHOLAR BENCH — every seat labeled as such. Human review invited, never waited for.**
**REVIEW AND COUNSEL ONLY — this bench edited nothing in the corpus, the ledger or any layer.**
Sitting: 2026-09-05, after `build/P41-COUNCIL/{council_900_charter,coverage_matrix_v1,preliminary_audit}.json` landed.
Scope read: the canonical ledger `framework/_FRAMEWORK-40000.json` (40,000 points), the P41 trio, `charter/*.md`,
`00-CORPUS/MANIFEST.sha256`, and sampled artifacts under `build/P9,P17,P20–P28,P30–P37,LENS-AUDIT,REPAIR-2` and `reviews/`.
This bench **documents, never adjudicates; never a mufti.** It issues no ruling and weighs no school against another.

---

## VERDICT IN ONE PARAGRAPH

**The waqf's inputs are sound; its outputs are half-sealed; its audit instrument is miscalibrated; and it has no
second copy.** Three things this bench re-derived independently and confirms: the corpus fixity is real (**195/195
files in `00-CORPUS/MANIFEST.sha256` recompute sha256-equal, 0 drift, 0 missing** ✅ VERIFIED), the build carries
**zero third-party dependencies** — stdlib Python only across 144 framework modules, the best long-horizon decision
in the whole engineering record — and the bidi discipline is genuine (`dir='rtl'` on 127 of 390 pages). Against that:
**`git remote -v` returns nothing** — 40,000 points, a 399 MB build tree, one branch, one Windows disk, under `…\Downloads\`, with
no off-site copy, which is a direct measured failure of Founding Principle 5 ("a stranger with one archive copy can
reconstruct all of it"). **P31 AL-MIZAN's own seal is lying about 15 of the 185 files it covers** and 69 more files
sit outside it, and **nothing watches it**; `verify_seals.py` itself **exits 1 today** (P30 drift) and is wired into
**no gate battery**, so the P41 preliminary audit's "batteries GREEN" is true and incomplete in the same breath.
**53.8% of the ledger's points (21,500) have no pillar seal at all.** And the coverage matrix that this council is
using as its map is **built by a probe that accepts any one of three loose substrings**: this bench reverses **27 of
the 36 PRESENT cells** in its own families, and changes **47 of 72 cells outright** — every one of the ten
engineering areas was marked PRESENT on the word "engineering" appearing three times, all three of them the phrase
*"the engineering bench"*. Worse, the matrix is
**not idempotent**: re-running `p41_council.py` today would read its own previous output (which lists all 313 area
names) and flip every ABSENT cell to TOUCHED.

---

## 0 · WHAT THIS BENCH INDEPENDENTLY RE-DERIVED (the all-clear, stated first)

A false alarm costs the trust a false all-clear costs. These were measured by this bench, from the artifacts, before
any finding was written.

| Claim | Bench 6's independent measure | Verdict |
|---|---|---|
| `00-CORPUS` inputs are sealed and intact | 195 manifest entries recomputed: **195 ok · 0 drift · 0 missing** | ✅ VERIFIED |
| The ledger holds 40,000 points, contiguous | `n` unique = 40,000; `min`=1 `max`=40000; `n == range(1,40001)` → **True** | ✅ VERIFIED |
| Every point carries pillar/domain/block/purpose | 40,000/40,000 on all four fields | ✅ VERIFIED |
| The build has no dependency rot exposure | import census, 144 modules: stdlib only (`io re collections hashlib glob statistics unicodedata difflib random time subprocess`) + local | ✅ VERIFIED |
| RTL text is direction-aware | `dir='rtl'` on **127** of 390 built pages; `.ar` blocks carry it | ✅ VERIFIED |
| Urdu typography is real, not a fallback | Nastaliq font stack specified on **126** pages | ✅ VERIFIED |
| The consensus shelf's honesty discipline holds | `build/P17-TAFSIR/D1/consensus_113_4.json` carries 6 `honesty_notes`, an empty-shard declaration, a named panel fix and a verbatim-substring proof | ✅ VERIFIED — exemplary |
| P36 bench packs are real, readable rows | 10 packs × 100 = **1,000 rows** on disk (+100 in P32 W-M) | ✅ VERIFIED |
| DOI is not falsely claimed | `datasets_layer.json` row 3: `"DOI": "AWAITED - owner-gated"` | ✅ VERIFIED — honestly declared |

---

## 1 · MATRIX CORRECTIONS — 72 AREAS RE-PROBED

**Why v1 needed correcting.** `framework/p41_council.py:probes()` takes the first three tokens longer than three
characters and marks an area PRESENT if **any one** of them appears as a **substring** anywhere in the ledger's
1,219,320-character domain text. It also reads `pillar + domain + purpose` but **not `title`** — so it over-fires on
generic words and under-fires on areas whose evidence lives in titles. Bench 6 re-probed all 72 of its areas with
**discriminating terms**, read the **actual ledger line** behind every hit, and excluded `build/P41-COUNCIL/*` and
`framework/p41_*.py` from the layer text (they contain all 313 area names verbatim — see **F7**).

**Net for Bench 6's families: v1 read PRESENT 36 · TOUCHED 22 · ABSENT 14. Corrected: PRESENT 13 · TOUCHED 29 · ABSENT 30.**
**47 of 72 cells (65%) change** — PRESENT→ABSENT **18** · PRESENT→TOUCHED **9** · TOUCHED→ABSENT **7** ·
TOUCHED→PRESENT **4** · ABSENT→TOUCHED **9**. **27 of v1's 36 PRESENT cells do not survive inspection.**

*This bench's own probe needed correcting twice, and says so:* its first pass marked **Superstition studies**
PRESENT on the token `omen` — which is the substring inside **w-omen** and phen-**omen**-on — and marked
**Cybersecurity** PRESENT on `adversar`, which in this repo is always "adversarial *panel*". Both were caught by
reading the matched line. That is the discipline F7 asks of the matrix: **a cell is not evidence until the line
behind it has been read.**

### 1a · Cells this bench REVERSES (v1 wrong)

| # | Area | v1 | v1's actual trigger | Bench 6 | Evidence |
|---|---|---|---|---|---|
| 1 | Civil engineering | PRESENT | `"civil"` → `n=1601 C2 · Mercy in **civil**ization` | **ABSENT** | no engineering domain in 40,000; only ledger "engineering" hits are `n=13991/16591/16891` = *"the **engineering bench** attacks"* |
| 2 | Mechanical engineering | PRESENT | `"mechanical"` → `n=15361 P30 · …made **mechanical**` | **ABSENT** | same three hits |
| 3 | Electrical engineering | PRESENT | `"engineering"` (3 hits, all "the engineering bench") | **ABSENT** | `circuit` / `voltage` searched in the engineering sense → absent |
| 4 | Chemical engineering | PRESENT | `"engineering"` | **ABSENT** | as above |
| 5 | Aerospace engineering | PRESENT | `"engineering"` | **ABSENT** | as above |
| 6 | Biomedical engineering | PRESENT | `"engineering"` | **ABSENT** | as above |
| 7 | Energy systems | PRESENT | `"systems"` → **one** hit: `n=20801 P31 · …kana/inna/zanna SISTER-**SYSTEMS**` | **ABSENT** | one grammatical hit carried four areas |
| 8 | Control systems | PRESENT | `"systems"` (same single hit) | **ABSENT** | "feedback loop" in ledger = `n=13401 P27 · Growth & stewardship` (a product feature) |
| 9 | Industrial & manufacturing eng. | PRESENT | `"engineering"` | **ABSENT** | as above |
| 10 | Architecture (incl. Islamic) | PRESENT | `"incl"` (4 hits) | **ABSENT** | `"architect"` in ledger = `n=61/161/261 · **Page architect**ure … mobile-first` — web layout, not the discipline |
| 11 | Machine learning | PRESENT | `"machine"` (149) → `…labelled **machine**-scholar` | **ABSENT** | `sklearn` 0 · `numpy` 0 · `torch` 0 imports; import census is stdlib-only |
| 12 | Computer vision | PRESENT | `"vision"` (200) → `n=5301 C6 · **Vision**, eternity & preparation` | **ABSENT** | 0 image pipeline; the 86 `ocr` file-hits are the substring in *"hy**pocr**ites"* |
| 13 | Data science & visualization | PRESENT | `"data"` (768) | **ABSENT** *(as a visual discipline)* | `<svg>` **0** · `<canvas>` **0** · "chart" in ledger = substring of **charter**; censuses are numeric-only (`LENS-AUDIT/E1_ocean_census.json`) |
| 14 | Database systems | PRESENT | `"systems"` (1 hit) | **TOUCHED** | no DBMS; JSON-file store. Real join discipline (`framework/stamp_joins.py`, `build/P31-MIZAN/W3-JOINS`) but no schema registry — see F12 |
| 15 | Distributed systems | PRESENT | `"systems"` (1 hit) | **TOUCHED** | `fleet/` + fsync append-only journal (`GRAPH_RULES` race-safety) are real; no distributed system exists |
| 16 | Human-computer interaction | PRESENT | `"human"` (365) → `…**human** review invited` | **TOUCHED** | 390 pages built; `aria-label` **0**, `role=` **0**, `alt=` **0** — see F13 |
| 17 | Applied cryptography | PRESENT | alias `"seal"` (198) — the ritual word | **TOUCHED** | the crypto is real but lives only in code (`hashlib.sha256`, 16 call sites) and 31 seal files; **no node governs it** |
| 18 | Comparative law (doc) | PRESENT | `"comparative"` → `n=5701 C7 · **Comparative** patience & trial` | **ABSENT** | 0 hits for common law / civil code / legal system |
| 19 | Development studies | PRESENT | `"development"` → `n=46 · how the discussion **develop**ed across centuries` | **ABSENT** | 0 hits for capacity building / development studies |
| 20 | International relations | PRESENT | `"relations"` (332) → `every **relation**ship carries its source` | **TOUCHED** | real material — `treaty` 44 files, `siyar` 5 — but no IR domain |
| 21 | Persian literature | PRESENT | `"literature"` → `the domain's own genre **literature**` | **TOUCHED** | the material is there but thin and incidental (`persian` 28 files, `farsi` 12); **`fa` = 1 of 74 translation editions and 0 of 29 tafsīr editions** — see F9 |
| 22 | Music theory (doc + positions) | PRESENT | `"documented"` (1,141 hits) | **TOUCHED** | `maqam` 15 files, `sama'` 6 — all incidental corpus text; no documented-positions layer |
| 23 | Sihr as documented phenomenon | PRESENT | `"documented"` (1,141 hits) | **TOUCHED** | genuinely curated at **exactly 3 verses** (`consensus_113_1/4/5.json`, incl. the Labid ibn al-A'sam knot-count divergence recorded without adjudication); 0 ledger nodes |
| 24 | Parapsychology claims | PRESENT | `"claims"` → `zero miracle **claims**, each item sourced` | **ABSENT** | 0 hits for parapsych / telepath / paranormal anywhere |
| 25 | Comparative mythology | PRESENT | `"comparative"` (200) | **TOUCHED** | one file: `build/P17-TAFSIR/D1/consensus_1_3.json` |
| 26 | Comparative religion | PRESENT | `"comparative"` (200) | **ABSENT** | 0 hits for comparative religion / ahl al-kitab / other faiths |
| 27 | World literature | PRESENT | `"literature"` → `the domain's own genre **literature**` | **ABSENT** | 0 hits for world literature / comparative literature |

*Subtotal 1a: **18 PRESENT→ABSENT · 9 PRESENT→TOUCHED = 27 of v1's 36 PRESENT cells reversed.***

### 1b · v1 TOUCHED cells that do not survive (7)

| # | Area | v1's trigger | Bench 6 | Evidence |
|---|---|---|---|---|
| 28 | Deep learning | `torch` | **ABSENT** | the 3 `torch` hits are the English word *torches* in `build/P36-BENCH/bench_pack_009.json`; 0 ML imports |
| 29 | Urban planning | — | **ABSENT** | 0 hits for urban planning / zoning / city plan outside P41 self-reference |
| 30 | Western jurisprudence theory (doc) | `"jurisprudence"` (21 files) — all meaning *fiqh* | **ABSENT** | 0 hits for legal positivism / natural law theory |
| 31 | Conventional finance (doc) | `"bond"` = the English word | **ABSENT** | 0 hits for interest rate / banking / conventional finance |
| 32 | World history | — | **ABSENT** | 0 hits after P41 exclusion (family E carries "Islamic civilization history" — not this cell) |
| 33 | Philosophy of history | — | **ABSENT** | 0 hits for philosophy of history / historiograph / periodization |
| 34 | Orientalism & its critique (doc) | — | **ABSENT** | 0 hits for orientalis / Wansbrough / Schacht / Goldziher / Nöldeke |

### 1c · Cells this bench RAISES — v1 under-read them (13)

| # | Area | v1 | Bench 6 | Evidence |
|---|---|---|---|---|
| 35 | Natural language processing | TOUCHED | **PRESENT** | v1's alias list missed it. Ledger P30/P31 carry morphology/lemma/i'rāb domains; `00-CORPUS/morphology.json` 15.1 MB sealed; treebank 139,376 tokens / 121 relations. The strongest O cell in the waqf |
| 36 | Computational linguistics | TOUCHED | **PRESENT** | P31 W4 dependency relations; `treebank` in 65 files; `Quranic.csv` sha-sealed at `framework/p31_w9_gates.py:22` |
| 37 | Management & organizational science | TOUCHED | **PRESENT** | four real ledger domains ≈ 400 points: P22 *Contributor & review workflows*, P22 *Recognition & governance channels*, P25 *Governance & succession*, P27 *Growth & stewardship* |
| 38 | Biblical studies (doc) | TOUCHED | **PRESENT** *(frame only)* | genuine ledger domain `n=7801 P9 · Tawrah, Zabur, **Injil** and the Suhuf` (300 points) and a real computed layer `build/P9-BOOKS/D1-REGISTER/books_register_layer.json` (Tawrah 16 refs, Injil 12, Zabur 3, Suhuf 8, Furqan 7). **But the layer is a Qur'anic occurrence census only** — no Biblical text, no comparison, no isrā'īliyyāt register |
| 39 | Cybersecurity | ABSENT | **TOUCHED** | tamper-evidence via 31 seal files is real; `threat model` **0**, `attack surface` **0**; the 206 `adversar` hits are all *"adversarial panel"* (review), not security — see F17 |
| 40 | Telecommunications | ABSENT | **TOUCHED** | genuine and deliberate: `n=75/175/275 · **Low-bandwidth pages** — tiny, zero-JS versions for slow connections`. v1 missed it; the concern is engineered, the discipline is not named |
| 41 | Anthropology | ABSENT | **TOUCHED** | `kinship` in 26 files (ansāb/nasab material in P8/P11) — anthropological substance without the discipline |
| 42 | Economics | ABSENT | **TOUCHED** | v1 read domain+purpose only and missed the title: `n=12001 P23 · Qur'an & medicine, **economics**, education…` — one title fragment inside a composite domain, no domain of its own |
| 43 | Criminology | ABSENT | **TOUCHED** | `penal` 29 files, `crime` 7, `hudud` in `build/P26-SCIENCES/D1-FIQH/fiqh_documented_layer.json` — documented, not analysed |
| 44 | Judaic studies (doc) | ABSENT | **TOUCHED** | `rabbin` 6 files, `isrā'īliyyāt` 1 (`build/P18-HADITH/D1/SOURCES.md`) — incidental, uncurated |
| 45 | Art history | ABSENT | **TOUCHED** | `ornament` 15 files, `illumination` 9 — all incidental corpus vocabulary; `calligraph` **0** outside P41. See F20 |
| 46 | Astrology & horoscopes (doc + kahānah rulings) | ABSENT | **TOUCHED** | `kāhin` 5 files, `astrolog` 3 — all raw `hadith_layer_dataset.json` text, never curated. See F8 |
| 47 | History of esotericism (doc) | ABSENT | **TOUCHED** | `esoteric` 4 files — incl. a real refusal line at `build/C2-BISMILLAH/D1-THE-NAME/SOURCES.md:111` *"Esoteric or numerological treatments of the Name — excluded as per point 907"* |

### 1d · Cells confirmed — v1 right, evidence restated (25 of 72)

**PRESENT (9 confirmed):** Algorithms & data structures (ledger titles carry *O(1)*, *index build*, *dedup*;
`framework/corpus_index.py` rebuilds in 14.5 s) · AI safety & alignment (P21-AI D1/D3, 200 points) · Knowledge
representation & ontologies (`charter/MASTER_ONTOLOGY.md`, `charter/GRAPH_RULES.md`, `build/P37-GRAPH/graph_layers.json`) ·
Information retrieval (`corpus_index.py`: 1.09 s over 68,921 reports) · Software engineering (Quality Ratchet L1–L10 in
`charter/THE_CONSTITUTION.md`; "selftest" in 427 files; batteries G18–G31) · Archival science (`archival-format print`
on 132 nodes; P25-D1 mirrors; provenance in 43 files) · Human geography (`C7 · Geography & chronology`,
`P14 · Narratives & geography`) · Urdu literature (9 Urdu tafsir editions held: tafhim_ur, bayan_ur, ahsan_ur,
usmani_ur, saadi_ur, ibnkathir_ur, zilal_ur, tazkir_ur, maarif) · Translation studies (P30 Wing C: registry of 74
editions, fidelity matrix, loss taxonomy, compensation registry).

**TOUCHED (11 confirmed):** Digital humanities (OpenITI, 33 files; no TEI) · Islamic economics & finance (zakāt 62 /
ribā 27 files as Qur'ānic vocabulary; no mu'āmalāt layer, and **no waqfiyyah instrument for the waqf itself** — F19) ·
Media & communication studies (P24-D3 Distribution, P22-D3 Launch) · Library & information science (per-wave manifests
are a real accession trail; `catalog` 4 files, no authority file — F22) · Political science (`statecraft` 5 files, in
`build/C2-BISMILLAH/D8-MERCY-CIVILIZATION/civilization_mercy_layer.json`) · Christian theology (`christian` 61 files,
`gospel` 41 — incidental corpus text) · Film & theatre studies (`film` 8 files, all corpus text) · Numerology & the
abjad heritage (**the refusal law is real but local**: `numerolog` in 26 files, and a numerology exclusion in **20 of
134** domain `SOURCES.md` — F8) · Palmistry & divination (`divination` 8, `azlām` 1) · Dream interpretation
(`ibn sirin` 9, `ru'yā` 4, `dream interpret` 2) · Folk medicine (`ḥijāma` 3, `cupping` 3, `build/C4-MUHAMMAD/D3-SHAMAIL/shamail_layer.json`).

**ABSENT (5 confirmed):** Robotics (0 hits of any kind) · Sociology (0) · Demography (`population` in the ledger =
*"Graph population — unit 1/10"*) · Museology (`exhibit` 2 files, both corpus text) · Superstition studies
(`tiyara` 0 · `'adwā` 0 · `superstit` 0 — v1 was right and this bench's first pass was wrong).

---

## 2 · THE FINDINGS — 25, WORST FIRST

Centre-clause key: **(a)** the Qur'ān as centre of research · **(b)** everything returns to Allah ﷻ ·
**(c)** meaning derived through sīrah/sunnah/ḥadīth · **POLISH** = touches none of the three.

| # | Finding | Area(s) | Cell | Sev | Clause | THE FILL |
|---|---|---|---|---|---|---|
| **F1** | **No second copy exists anywhere.** `git remote -v` → **empty**; the remote count is **0**; 1 branch (`master`), 5 tags (`math-v1, v0, waqf-complete-13800, waqf-complete-17400, waqf-complete-40000`); the `10-WAQF` subtree alone measures **399 MB**, on one Windows disk at `C:\Users\walee\Downloads\Quran Research\`. Founding Principle 5 requires that "a stranger with **one archive copy** can reconstruct all of it"; there is no copy to be a stranger with. `build/P25-PERMANENCE/D1-MIRRORS/permanence_assets_layer.json` names "the git chain" and "the v1 frozen mirror (= git tag `math-v1`)" as the permanence assets — **both live on the same disk as the thing they protect**. A single disk failure, a single ransomware event, or Windows Storage Sense reaching a `Downloads` subtree ends the waqf. | Archival science · Distributed systems · Library & IS | PRESENT | **HIGH** | (a)(b)(c) — total-loss risk to the whole | **One acquisition + one engine, today.** (i) push the repo to ≥2 independent remotes in ≥2 jurisdictions; (ii) a `mirror_check.py` engine that asserts ≥3 copies exist and fails the battery if not; (iii) one deposit of the sealed corpus + ledger to a permanent archive (the Zenodo path is already proven per `datasets_layer.json`). ~40 nodes under P25-PERMANENCE. **Owner decision + machine executes.** |
| **F2** | **P31 AL-MIZAN's seal is lying.** `build/P31-MIZAN/W9-C10/SEAL.json` records sha256 for **185** files. This bench recomputed all 185: **15 DRIFTED** (8.1%) — `W5-T1/ma_types_layer.json`, `W5-T3/particle_homographs_layer.json`, `W7-M1/strata_layer.json`, `W8-V5/tvf_v1.json`, `W8-V10/coverage_proof.json`, `W9-GATES/gate_battery.json`, `W9-GATES/gate_battery2.json`, `PAGES/index.html`, `PAGES/W9-D10--grand_certification.html` + 6 more. **254 files exist on disk; 69 are outside the manifest entirely.** 14 commits touched `build/P31-MIZAN` after the seal was cut (`2aef579`), and it was never re-cut — against the waqf's own `PW3-F1` policy ("seal-what-you-read … **re-stamp at re-seal**", commit `32010d9`). P31 is **9,000 ledger points — 22.5% of the waqf**. | Applied cryptography · Software eng. · Archival science | TOUCHED | **HIGH** | (a) | **An engine + a gate.** `verify_manifests.py` that walks every `SEAL.json`-class manifest (not only `_*-PILLAR-SEAL.sha256`), reports drift and coverage, and **exits non-zero**; then re-cut P31's seal over all 254 files with a dated drift note. Wire into G32. **Machine.** |
| **F3** | **The drift-watch is RED and no battery calls it.** `python framework/verify_seals.py` today prints `DRIFT P30: 1 changed` (`D-D6-BENCH/bench_charter.json`) and **exits 1** ✅ VERIFIED. Yet `preliminary_audit.json.batteries` reports all three batteries GREEN — because `verify_seals` is called by **none** of them (`grep -n "verify_seals" p31_w9_gates.py p31_w9_gates2.py p40_gates.py` → 0 hits). `build/P25-PERMANENCE/D3-RITUALS/rituals_layer.json` declares the ritual "*every `process_landings` sweep recomputes all seal hashes; drift is named the hour it happens*" — the hour has not been named. | Software eng. · Archival science · Cybersecurity | PRESENT | **HIGH** | (a) | **Two lines of wiring.** Add `verify_seals.py` and the new `verify_manifests.py` to the gate battery as G32/G33; make the P41 prelim report their exit codes, not just the three it currently runs. Then close the P30 drift with a deliberate re-seal note. **Machine.** |
| **F4** | **53.8% of the ledger has no pillar seal at all.** 31 `_*-PILLAR-SEAL.sha256` files exist, covering C1–C8, P8–P28. **No seal file exists** for `P31-MIZAN` (9,000 pts), `P32-SHAHADAT` (1,700), `P33-RIJAL` (2,000), `P34-OCEAN` (1,500), `P35-EDITIONS` (2,500), `P36-BENCH` (1,000), `P37-GRAPH` (1,500), `REPAIR-2` (800), `P28-SIRAH-X` (1,500) — **21,500 points**, plus `LENS-AUDIT`, `C5-JIBRIL-ANGELS` and `P41-COUNCIL`. These are the *newest and largest* layers: the seal discipline was strongest when the waqf was smallest. | Archival science · Applied cryptography | PRESENT | **HIGH** | (a) | **Run the existing ritual.** `framework/close_pillar.py` already cuts seals; run it over the 12 unsealed trees, then let G32 assert *every* build directory carries a seal. ~12 nodes. **Machine.** |
| **F5** | **The gap register's own prescribed fill would corrupt verbatim corpus text.** `p41_gaps.py:43` prescribes "rerun `adab_scan` over `build/P31..P37`". This bench ran `adab_scan.sweep()` over those trees: **4,456 flags across 14 files** — field breakdown: `hadith_en` **3,289** · `verse_en` **1,155** · `gist_en` **10** · `_meta.law` **2**. **4,444 of them (99.7%) are in `hadith_en` and `verse_en`** — corpus-verbatim English translations of ḥadīth and āyāt in `build/P36-BENCH/bench_pack_00*.json` and `build/P32-SHAHADAT/W-M/bench_pack_001.json`. `adab_scan.EXEMPT` exempts `en`, `quote_en`, `held_en`, `ar` — **but not `hadith_en` or `verse_en`**, because it is a hand-maintained literal field-name allow-list that the newer layers outgrew. **If the prescribed fill is run with `exalt()` as the older factories do (`c1_factory.py:206`, `consensus_tool.py:123`), it will insert ﷻ into 4,444 transmitted translation strings** — a verbatim-law violation at scale, executed in the name of the adab law. | Software eng. · Computational linguistics · Data/DB | PRESENT | **HIGH** | (c) — the verbatim law protects transmitted text | **Fix the engine before running the fill.** (i) invert the allow-list: mark waqf-authored fields explicitly (`_authored: [...]` in each layer's `_meta`) and sweep **only** those; (ii) a `--prove-fail` test that asserts `exalt()` leaves every corpus-verbatim field byte-identical; (iii) *then* run the sweep. **Machine.** |
| **F6** | **The three load-bearing doctrinal laws have no executable gate.** (i) **never-a-mufti**: `build/P21-AI/D1-NEVER-A-MUFTI/never_a_mufti_layer.json` gives its verifier as `"every engine's law line"` — a prose convention, not a test; `grep -rn "mufti" framework/*.py` → 5 hits, **all of them the phrase itself, none a check**. (ii) **"No claim of miracle from arithmetic, ever"** (FP4 + `THE_CONSTITUTION.md` red lines): **zero** detectors anywhere. (iii) **the Adab Law**: `adab_scan.RX_FLAG` tests exactly one clause (`Allah` → ﷻ); ﷺ after the Messenger, عليه السلام after prophets, رضي الله عنه/عنها after ṣaḥābah, رحمه الله after scholars are **unchecked by any engine**. The batteries G18–G31 are all structural (staleness, totals, id-joins, dimension counts) — not one is doctrinal. | AI safety & alignment · Software eng. | PRESENT | **HIGH** | (b) | **One engine, three detectors.** `doctrine_gates.py`: a ruling-language detector over every published string (imperative fiqh verbs + "you must/it is obligatory/the ruling is"), an arithmetic-miracle claim detector (numeric claim + miracle/proof vocabulary in one sentence), and an honorific-completeness detector covering all five clauses with the same field-awareness fix from F5. Gate as G34–G36. ~60 nodes under P21-AI. **Machine builds, owner ratifies the word-lists.** |
| **F7** | **The coverage matrix — the council's own map — is miscalibrated, and it is not idempotent.** `p41_council.py:probes()` returns the first three tokens >3 chars and marks PRESENT on **any** substring hit. Measured consequences in Bench 6's families alone: ten engineering areas PRESENT on `"engineering"` (3 ledger hits, all "the engineering bench"); four areas PRESENT on `"systems"` (**one** hit — "SISTER-SYSTEMS"); Music theory and Sihr PRESENT on `"documented"` (1,141 hits); Computer vision on `"vision"` (C6 eschatology); Machine learning on `"machine"` ("machine-scholar"); Parapsychology on `"claims"`. **27 of v1's 36 PRESENT cells reversed; 47 of 72 cells changed (65%).** Separately: the layer-text glob `build/*/*.json` now matches `build/P41-COUNCIL/coverage_matrix_v1.json` (9,572 bytes, under the 400 KB cap), **which lists all 313 area names verbatim** — so a re-run of `p41_council.py` today flips **every one of the 88 ABSENT cells to TOUCHED**. The instrument reads its own output. | Information retrieval · Data science · Software eng. · Research method | PRESENT | **HIGH** | (a) — the audit's own evidence standard | **Rebuild the probe as v2.** (i) require **all** discriminating terms, not any; (ii) probe `title` as well as domain+purpose; (iii) word-boundary regex, never bare substring; (iv) exclude `build/P41-COUNCIL/**` and `framework/p41_*.py` from the layer text; (v) emit the **matched ledger line** as evidence for every PRESENT cell so a cell can be falsified by reading it; (vi) assert idempotency (run twice, diff = ∅). **Machine.** |
| **F8** | **The whole contested-knowledge family is unregistered — 0 of 40,000 nodes.** 22 discriminating terms probed against the ledger (`abjad, numerolog, astrolog, kahin, kahanah, sihr, magic, ruqya, amulet, tamima, dream, ruya, tabir, superstit, tiyara, esoteric, batini, mythol, folk, parapsych, azlam, horoscope`) → **zero hits**. Yet the Book raises these subjects directly (2:102, 5:90, 12 — Yūsuf's ru'yā, 52:29, 69:42, 113:4) and the material is already in the held ocean (`sihr` 4 files, `kahin` 5, `ibn sirin` 9, `divination` 8 — all incidental hits inside raw `hadith_layer_dataset.json`, uncurated). Only **three verses** are curated (`consensus_113_1/4/5.json`). The refusal law **does** exist at charter level — FP4 "No miracle is ever claimed from arithmetic" and `THE_CONSTITUTION.md` red lines — but it reaches only **20 of 134** domain `SOURCES.md` files and **no ledger node**. | all 10 S-family areas · Comparative religion | ABSENT/TOUCHED | **HIGH** | (a) + (c) | **One domain cluster, documented-never-practiced.** ~250 nodes: (i) the Qur'ānic register of sihr/kahānah/azlām/ru'yā/tiyarah, located mechanically via `corpus_index`; (ii) the tradition's own verdicts quoted verbatim from the held graded corpus, carriers named, **no adjudication**; (iii) the abjad heritage documented beside FP4 as the waqf's stated refusal, promoted from `SOURCES.md` prose to a ledger law with the F6 detector behind it. The register already scopes this at MEDIUM — **this bench raises it to HIGH** and supplies the measure. **Machine builds, scholars verify.** |
| **F9** | **The meaning shelf speaks three languages; the translation shelf speaks fifty-five.** `00-CORPUS/tafsir.json` holds **29 tafsīr editions: Arabic 12 · English 8 · Urdu 9. Zero Persian, zero Turkish, zero Indonesian/Malay, zero Bengali, zero Hausa, zero French, zero Swahili.** `build/P35-EDITIONS/editions_inventory.json` profiles 74 translations in **55 languages** — but `fa` = **1** edition, and the *meaning* layer (where the centre test lives) has none. The entire Persian tafsīr branch — Maybudī's *Kashf al-Asrār*, Abū al-Futūḥ Rāzī's *Rawḍ al-Jinān*, *Tafsīr-e Sūrābādī* — a thousand-year tradition, is absent. Measured against readership: the largest Muslim populations on earth (Indonesia, Bangladesh, Turkey, Iran, Nigeria) have **no** tafsīr edition in the meaning shelf. | Persian literature · World literature · Translation studies · Human geography · Demography | PRESENT | **HIGH** | (c) | **A targeted acquisition wave.** Priority order by readership × openness: Persian (Maybudī, PD; *Nemūneh* where licensable), Turkish (Elmalılı Hamdi Yazır, PD 1942), Indonesian (Kemenag tafsir), Bengali, Hausa. Each landed under the existing law: licence recorded, sha-sealed, tier-stamped. ~150 nodes under P35/P17. **Machine acquires; native reviewers verify (HUMAN SEAT: one native reader per language).** |
| **F10** | **1,100 review rows prepared, 0 consumed, and no channel to consume them.** `build/P36-BENCH/bench_pack_002…011.json` = 10 packs × 100 rows; `build/P32-SHAHADAT/W-M/bench_pack_001.json` = 100 more. `build/P36-BENCH/verdict_intake.json` → `"verdicts": []` — **empty** ✅ VERIFIED, matching `preliminary_audit.other_measures.human_verdicts: 0`. The intake schema is excellent (scholar · school · verdict · establishes · divergence — and *both sides where schools differ*). What does not exist: a **named addressee**, a delivery channel, a capacity estimate, or a turnaround expectation. At 5 minutes per row the standing queue is ≈92 scholar-hours before the first pack clears — a number nobody has been shown. | Management & org. science · Media & communication studies | PRESENT | **HIGH** | (c) | **Not more machine work — a distribution act.** (i) one page listing the packs with a one-click intake form matching the schema; (ii) a named first reviewer and a single pack (001), one afternoon; (iii) a capacity model in the register so the queue depth is visible. The register already rates this CRITICAL. **OWNER + HUMAN SCHOLARS — nothing the machine can do substitutes.** |
| **F11** | **The canonical ledger's own header is wrong about the ledger.** `_FRAMEWORK-40000.json._meta.total = 26400` while `len(points) = 40000` ✅ VERIFIED. `_meta.version = "2.0.0"` while `meta.v4.version = "4.0.0"` — **two live version fields in one file, disagreeing**, with `_meta` carrying six historical `vN.N.N` keys alongside. Any consumer that trusts the declared total under-reads the waqf by **13,600 points (34%)**. | Database systems · Archival science · KR & ontologies | PRESENT | **MED** | (a) | **One append-only correction + one gate.** Set `_meta.total` from `len(points)` at write time; collapse to a single authoritative `version`; G37 asserts `_meta.total == len(points)` and fails otherwise. ~4 nodes. **Machine.** |
| **F12** | **Schema drift inside the canonical ledger.** Points **26,401–40,000 (13,600 of 40,000, 34%)** carry only `n, pillar, domain, block, purpose`. They are **missing `kind`, `pillar_name` and `title`** — which the first 26,400 all carry ✅ VERIFIED. Any join, page render or coverage probe keyed on `title` silently drops a third of the waqf; this is *precisely* why the v1 matrix probe (which reads `domain + purpose` only) and this bench's first probe disagreed on Architecture and Biblical studies. There is no schema validator: `charter/MASTER_ONTOLOGY.md` is **3,136 bytes** and `charter/GRAPH_RULES.md` **1,833 bytes**, governing a 40,000-node ledger by prose alone — Quality Ratchet **L2** ("one shared schema module per layer") is asserted, not enforced. | KR & ontologies · Database systems · Software eng. | PRESENT | **MED** | (a) | **A machine-readable schema + a validator.** JSON-Schema for the point record; backfill `kind`/`pillar_name`/`title` for 26,401–40,000 from their pillar/block (append-only field add); G38 validates all 40,000 against the schema every battery run. ~20 nodes. **Machine.** |
| **F13** | **Zero accessibility markup in a pillar named "Education & accessibility."** Across **390** built HTML pages: `aria-label` **0** · `role=` **0** · `alt=` **0** ✅ VERIFIED. **0 pages** tag Arabic or Urdu spans with `lang=`, while **143** declare `<html lang='en'>` and then render Arabic and Urdu inside it — a screen reader voices the Qur'ān in an English voice; hyphenation and font selection fall back to English. Direction is handled well (`dir='rtl'` on 127 pages) — the failure is language identity, not direction. `build/P24-ACCESS/D2-EDUCATION/education_access_layer.json` (100 ledger points) claims "print accessibility … semantic headings; **dir-aware text**; high contrast (**declared per page**)" with `verify: "any pages_trilingual.html"` — a claim whose verifier is "go look". | Human-computer interaction · Media studies | TOUCHED | **MED** | (a) — access to the Book | **A template fix + a gate.** Add `lang='ar'` / `lang='ur'` to every `.ar` / `.ur` block and `alt`/`role`/skip-links to the page template (one template, 390 pages regenerate); `a11y_gate.py` asserts per page: root lang present, every RTL block language-tagged, contrast ratio ≥ 4.5:1, zero untagged images. Gate as G39. ~25 nodes under P24. **Machine.** |
| **F14** | **The thinnest evidence band in the waqf: P20–P27.** Eight branch pillars = **2,400 ledger points (6%)**, 24 domains. Their substantive layers total **60 domain-layer rows + 81 Qur'ān rows** — **≈2.5 substantive rows per 100 points**. Concretely: `build/P21-AI/D3-GUARDRAILS/guardrails_layer.json` carries **3 rows** for a 100-point domain, and all three merely *name existing scripts* (`prepanel_gate.py`, `reviews/`, `adab_scan.py`). `P26-SCIENCES` domains carry **2 rows each**. This is the "batch evidence" the prelim flagged, at its sharpest — and unlike P30/P31, these pillars have no dense measured layer behind them. To their credit the ḥadīth substrate is honest: `hadith_layer_dataset.json._meta.method` states plainly that "a term match can be incidental … this is the evidence substrate, **not** a curated thematic selection." | Research method · Management & org. · Software eng. | PRESENT | **MED** | (a) | **Per-node artifact pointers, worst band first.** Extend the tracker so every P20–P27 node names the **file** its claim lives in, not the run that built it; where no file exists, the node is re-marked honestly. Then decide per domain: fill or declare thin. ~2,400 re-marks. **Machine.** |
| **F15** | **Twelve true adab defects — ten of them in a file whose own `_meta` declares the adab law satisfied.** `build/P30-SIGHAH/D-B2-LANDING/mawquf_SG3.json` declares `"adab_law": "Allah ﷻ, Muhammad ﷺ, prophets 'alayhim as-salam honored…"` — yet **10 `gist_en` strings** (waqf-authored prose, prefixed `[gist]`, **not** corpus-verbatim) write the Majestic Name bare: `rows[0]` "begin with the remembrance of **Allah**, Lord of…", `rows[16/17/28/146]` "all praise belongs to **Allah** alone". `rows[18]` in the same file writes "Allah's ﷻ help" correctly — so the law is known and inconsistently applied. Two more sit in `_meta.law` strings, the layers' own law lines: `build/P31-MIZAN/W1-L8/muqattaat_layer.json` ("…oldest recorded position (**Allah** knows best…") and `build/P31-MIZAN/W4-S6/idafa_layer.json` ("ism→**Allah**→…"). These 12 are **true positives**, distinct from the 4,444 false positives of F5. | Software eng. · Computational linguistics | PRESENT | **MED** | (b) | Twelve string edits, then the F5-corrected sweep over all authored fields, then G36 (F6) so a `_meta` claim can never again outrun the file's contents. **Machine.** |
| **F16** | **The permanence pillar's self-description is stale in the direction that flatters it.** `permanence_assets_layer.json` row 1: "the pillar seals — **11** SHA256 seal files, drift-watched by `verify_seals.py` every sweep." Measured: **31** seal files matching `build/*/_*-PILLAR-SEAL.sha256`, and the drift-watch runs in **no** battery and **exits 1 today** (F3). The count understates by 20 — but "drift-watched every sweep" overstates by the whole mechanism. The layer's own law reads "every row names a REAL artifact **verifiable on disk this minute**"; row 1 is not. | Archival science · Software eng. | PRESENT | **MED** | (a) | Regenerate the row from a live `glob` + the actual exit code of `verify_seals.py`, so the claim cannot drift from the count again. 1 node. **Machine.** |
| **F17** | **No threat model exists.** the phrases `threat model` and `attack surface` searched case-insensitively across `framework/ build/ charter/ reviews/` → **0 hits each** ✅ VERIFIED. The 206 files matching `adversar` are all "adversarial **panel**" — review, not security. The waqf plans public community contribution (P22 *Contributor & review workflows*, 100 points) and a public API (P15-D3, P20-D3) over a corpus whose integrity claim **is** its value. There is no recorded adversary, no consideration of what a hostile contributor could insert into a registry, and no signature over the seals (a seal file on the same disk as its artifacts is tamper-**evident only against accident**, not against an editor). | Cybersecurity · Applied cryptography | TOUCHED | **MED** | POLISH | **One page, then one key.** (i) `THREAT-MODEL.md`: assets, adversaries (accidental editor, hostile contributor, disk failure, silent bit-rot), and what each control actually stops; (ii) sign the seal manifests (minisign/GPG) so integrity survives a writable disk; (iii) F1's off-site copies make bit-rot recoverable, not just detectable. ~15 nodes under P25. **Machine drafts, owner ratifies.** |
| **F18** | **The share-alike/non-commercial obligation is unstated waqf-wide.** `CC BY-NC-SA` (the OpenITI acquisitions) appears **170 times across 16 files**; `CC BY 4.0` in **3**. Those acquisitions are load-bearing: the P33 rijāl registries (6,143 text-blocks) and part of the P34 ocean derive from them. NC + SA propagates to derived layers — which sits in tension with FP5's promise that "any community, in any century, can carry it", and blocks several permanent-archive deposit terms that F1 would otherwise use. Per-wave licence records exist and are honest; **no single artifact states the resulting obligations.** | Comparative law (doc) · Archival science · Library & IS | TOUCHED | **MED** | POLISH | `LICENCE-LEDGER.md`: per source → its licence → which layers derive from it → what the waqf may do, must do, and cannot do; flag every layer whose licence would block archival deposit **before** F1 runs. The register already carries this at MEDIUM. ~10 nodes. **Machine drafts, owner ratifies.** |
| **F19** | **The waqf has no waqf instrument.** `governance_layer.json` states "governance is written instruments, not persons" and lists five charter files totalling **13,813 bytes**. None of them is a **waqfiyyah**: there is no named *nāẓir*/mutawallī, no succession trigger, no asset schedule, no dissolution clause, no jurisdiction. `P25 · Governance & succession` is 100 ledger points and `P22 · Recognition & governance channels` another 100; both describe *workflows*, not *title*. FP5 says "the founder's name may disappear" — today, if it did, nothing names who holds the trust. | Islamic economics & finance · Management & org. · Comparative law | TOUCHED | **MED** | (b) | A drafted waqfiyyah as a documented instrument (never a ruling): asset schedule, nāẓir succession, the conditions the founder sets, and the pointer to qualified scholars for its ṣiḥḥah. ~20 nodes under P25-D2. **Machine drafts the skeleton; HUMAN SEAT: a scholar of awqāf law + a lawyer in the chosen jurisdiction.** |
| **F20** | **The mushaf layer is text-only — no image, no plate, no folio.** 0 OCR, 0 image pipeline, 0 manuscript scan (F-cell 12 above). P35 profiles 74 editions by **character counts and coverage**; codicology, illumination, rasm variation as *written*, and the manuscript witnesses (Ṣanʿāʾ, Topkapı, Birmingham) exist in the waqf only as words about texts. The Art history seat notes the corollary: `illumination` appears in 9 files, `ornament` in 15 — all incidental corpus vocabulary; there is no register of mushaf art, and `calligraph` returns nothing outside P41. | Art history · Computer vision · Digital humanities | TOUCHED | **MED** | (a) | **An acquisition, not an engine.** Land open-licensed mushaf facsimile plates (Birmingham/Mingana CC, Tübingen, Corpus Coranicum) with licence + sha seal; a folio register keyed to sūrah/āyah; **no OCR claim and no palaeographic adjudication** — images placed beside the sealed text, nothing more. ~80 nodes under P35. **Machine acquires; HUMAN SEAT: a codicologist for any dating claim.** |
| **F21** | **Provenance digests truncated to 64 bits.** `hexdigest()[:16]` at `framework/c1_factory.py:266`, `framework/c3_layers.py:19`, `framework/corpus_index.py:42`, `framework/stamp_joins.py:26` — these 16-hex stamps are what `gates_audit_record.json` calls `source_hash_staleness_stamps`. For staleness detection against accident this is correct and cheap. For a stated 10,000-year tamper-evidence posture it is below the bar: a 64-bit digest has a ~2³² birthday bound and is not a provenance guarantee against an editor. The full-length seals (`verify_seals.py`, `SEAL.json`) are correct — the inconsistency is that both are described with the same word, "seal". | Applied cryptography | TOUCHED | **LOW** | POLISH | Keep 64-bit for staleness but **rename the field** (`staleness_stamp`, not `sha256`) so nothing reads it as provenance; use full digests wherever the word "seal" is used. Note the single-algorithm dependency in `THREAT-MODEL.md` (F17). ~4 nodes. **Machine.** |
| **F22** | **No accession register and no persistent identifier.** The waqf holds 74 translation editions, 29 tafsīr editions, 5 rijāl works, two ḥadīth oceans and a treebank, tracked by per-wave manifests. There is no single accession register (one row per acquired work: date, source, licence, sha, which layers consume it) and **no DOI/ARK/Handle**. Stated honestly, to the build's credit: `datasets_layer.json` row 3 reads `"DOI": "AWAITED - owner-gated"`. Without a persistent identifier nothing outside this disk can cite the waqf, which is the same failure as F1 seen from the reader's side. | Library & information science · Archival science | TOUCHED | **LOW** | POLISH | One `ACCESSIONS.md` generated from the existing manifests (zero new information required — it already exists, scattered); mint the DOI when F1's deposit lands. ~8 nodes under P20-D2. **Machine + owner's GO for the deposit.** |
| **F23** | **Device censuses cover 4% of the shelf — declared, not hidden.** `editions_inventory.json`: **3 of 74 editions** device-censused, English only. The layer's own law states the reason and refuses to guess: "*the per-language marker programs are earned by native review — registered, never guessed*", plus a known-positive law that **refuses** an English census run without its 55:13 anchor. This bench records it as a **model of honest scope declaration** and sizes the remainder rather than faulting it: **71 editions (96%) uncensused**. | Translation studies · Data science | PRESENT | **LOW** | (c) | Exactly what the layer says: native-reviewer slots, one language at a time, each earning its marker set. **HUMAN SEATS: one native reader per language.** |
| **F24** | **Two engineering strengths this bench asks be protected as laws, not habits.** (i) **Zero third-party dependencies** — the import census across 144 framework modules returns stdlib only; there is no lockfile because there is nothing to lock. For a 10,000-year posture this outperforms every other decision in the build, and it is currently a habit that one `pip install` could end. (ii) **Input fixity is perfect** — 195/195 corpus files verified this sitting. Both are undefended: nothing fails if a dependency is added, and `MANIFEST.sha256` is verified by no battery either. | Software eng. · Archival permanence | PRESENT | **LOW** | POLISH | Two asserts: G40 fails if any `import` outside the stdlib+local allow-list appears in `framework/`; G41 runs the `00-CORPUS/MANIFEST.sha256` verification every battery run (this bench's run took seconds). ~4 nodes. **Machine.** |
| **F25** | **The two-axis claim, from the data-modelling vantage.** Confirming the prelim's measure independently: **0 of 40,000** points carry `axis_existence` or `axis_evidence`; the declaration holds only by pillar inheritance. From the KR seat the concern is narrower than "add a field": there is **no schema** in which such a field could be validated (F12), so stamping axes now would produce 40,000 unvalidated strings. The order matters — schema first, then stamps. | KR & ontologies · Database systems | PRESENT | **LOW** | (a) | **Sequence the register's existing GAP-002 behind F12.** Ship the JSON-Schema, then derive both axis stamps mechanically from pillar+block, then G42 asserts 40,000/40,000. **Machine.** |

**Nothing above is marked OPINION: every one of the 25 findings names a file, a command or a count that this bench ran.**
Two judgements about the *world outside the repo* are flagged inline as inference rather than measurement — that Windows
`Downloads` subtrees are exposed to automated cleanup (F1), and the readership figures behind the language-geography
argument (F9); the repo-side measures in both findings stand on their own without them.

---

## 3 · THE CENTRE TEST, FROM THIS BENCH'S VANTAGE

The prelim measured the central gap correctly and this bench does not dispute it: evidence layers span 59.4–80.0% of
the Book, meaning layers span 1.4–2.3%. Bench 6 adds one observation from the **Translation studies / Persian
literature / Demography** seats that sharpens rather than contradicts it: **the 2.3% is also monolingual in its
tradition-reach.** The consensus shelf's 146 verses are built from 29 tafsīr editions in **three** languages, and the
Persian, Turkish, Malay and Bengali commentary traditions contribute **zero** — so the "meaning derived through
sīrah/sunnah/ḥadīth" pass, when it scales from 146 verses to 6,236, will scale a shelf that has never read a Persian
mufassir. The acquisition (F9) is cheaper before the derivation pass than after it.

Of the 25 findings: **11 touch clause (a)**, **3 touch clause (b)**, **4 touch clause (c)**, **1 touches all three**
(F1), and **6 are POLISH**. The clause-(a) concentration is not about the Qur'ān's content — it is about *provenance*:
eleven separate ways in which a reader today cannot mechanically prove that the layers around the Book are what they
claim to be. F1–F4 are one finding wearing four coats, and closing them closes most of the bench's docket.

---

## 4 · THE FILL, CONSOLIDATED

| Kind | What | Findings | Size |
|---|---|---|---|
| **Owner decision (today)** | ≥2 off-site remotes + 1 permanent-archive deposit | F1 | one afternoon |
| **Engines** | `verify_manifests.py` · `doctrine_gates.py` (3 detectors) · `a11y_gate.py` · matrix probe v2 · JSON-Schema validator | F2,F3,F6,F7,F12,F13 | 5 engines |
| **Gate wiring** | G32–G42 (seals, manifests, doctrine ×3, a11y, schema, `_meta.total`, stdlib-only, corpus fixity, axes) | F3,F4,F6,F11,F12,F13,F24,F25 | 11 asserts |
| **Ritual re-runs** | `close_pillar.py` over 12 unsealed trees; P31 re-seal over 254 files; P30 drift closed | F2,F4 | one run |
| **Nodes** | contested-knowledge cluster ~250 (P43) · mushaf plates ~80 · language acquisitions ~150 · waqfiyyah ~20 · threat model ~15 · misc ~65 | F8,F9,F17,F19,F20 | **≈580 nodes** |
| **Acquisitions** | Persian/Turkish/Indonesian/Bengali/Hausa tafsīr; open mushaf facsimiles | F9,F20 | 2 waves |
| **HUMAN SEATS** | first bench sitting (any qualified scholar, pack 001) · one native reader per language · a codicologist for any dating claim · a scholar of awqāf law + a jurisdiction lawyer for the waqfiyyah | F10,F9,F20,F19 | **4 seat types** |

**Engine work before nodes.** F5 is the reason: the register's own prescribed fill, executed today, damages the corpus.
Fix the sweep, then sweep.

---

**BENCH 6 VERDICT: SEAL WITH GAPS** — the corpus inputs, the dependency posture and the honesty discipline are genuinely sound and this bench re-derived them itself, but the waqf currently has **no second copy** (F1), a **seal that lies about 15 files and covers only 46% of the ledger** (F2–F4), **no executable gate behind its three load-bearing doctrinal laws** (F6), and a **coverage matrix that reverses on inspection and re-reads its own output** (F7) — none of which touch the Book's content, and all of which must close before the seal means what it says.

---
*MACHINE-SCHOLAR. Documented, never adjudicated. Never a muftī. Human review invited on every line of this page and waited for on none.*