# THE FULL AUDIT — BENCH 5

**MACHINE-SCHOLAR BENCH. Every seat on this bench is a machine-scholar seat and is labeled as such.
No seat here adjudicates between schools. No seat here issues a ruling. No seat here is a mufti.
This bench measures, and reports what the measurement said.**

| | |
|---|---|
| **Bench** | 5 of the Council of 900 |
| **Families** | **L · Biology & life** (32) · **M · Medicine & health** (40) · **N · Mind, behavior & education** (20) — 46 areas, 92 seats |
| **Convened for** | THE FULL AUDIT of every band of the 61,600-node waqf, before the 41,600-node gift layer |
| **Sat** | 2026-09-05 |
| **Bands audited** | **B1** 1–13,800 · **B2** 13,801–17,400 (P30) · **B3** 17,401–26,400 (P31) · **B4** 26,401–40,000 (P32–P37 + allocations) · **B5** 40,001–61,600 (P42) |
| **Read first** | `build/P42-GAPFILL/full_audit_pointer_sweep.json` · `framework/_FRAMEWORK-61600.json` meta · `build/P41-COUNCIL/council_900_charter.json` |
| **Sources opened** | `framework/_TRACKER-JOURNAL.jsonl` (130,131 lines) · the whole-tree file index (3,832 files / 2,541 basenames) · `00-CORPUS/morphology.json` · `00-CORPUS/quran.json` · the four sealed hadith corpora (119,261 rows) · 130 `education_pack.json` · P30/P31/P32/P28-X/P34/P42 layer artifacts |
| **Prior sittings by this bench** | `reviews/COUNCIL-900-BENCH-5.md` · `reviews/COUNCIL-900-P42-BENCH-5.md` (this audit **corrects one of their claims** — §5) |

---

## 0 · THE VERDICT IN ONE PARAGRAPH

**The arithmetic is sound. The teaching is not.** Nine numbers re-derived across all five bands
reproduced **exactly, 9 for 9** — including two set-equality checks and two independent recounts
from per-row containers, listed with their commands in §1. Where the waqf counts, it counts truly.
But the gift layer is *public*, and the public surface of this waqf is its 130 education packs — the
one artifact class squarely inside this bench's family N. There, in band 1, the audit found
**19 children-layer statements, across 11 packs, that are false**, each self-stamped
`"true_sourced": true`, under a pack law that reads *"every item TRUE and sourced to the sealed
layer; no myth, no invention."* The Shamail pack tells a child *"manner - 647 reports appears in the
Book 647 time(s)"* — the term it names carries **ﷺ, which occurs 0 times in the Qur'an text**, and
647 is a count of *hadith reports*, correctly derived (`shamail_layer.json`: 647+116+122+49 = 934 = its
row count, verified). The scholarship underneath is right; the template that renders it to the public
is wrong, and it certifies itself. Alongside this: **423 of 984 glossary rows (43.0%) carry a
placeholder verse-reference**, and P42's 21,600 newest nodes rest their entire education block on
**24 files averaging 722 bytes**. Band 1 cannot carry a public gift layer until that template is
rebuilt. Bands 2–5 can, with the notes in §4.

---

## 1 · THE RE-DERIVATIONS — nine numbers, five bands, 9/9 exact

The task asked for five across four bands. Every row below was recomputed from the **source**, not
read from the header that asserts it. ✅ VERIFIED throughout.

| # | Band | Claim under test | Source re-run | Result |
|---|---|---|---|---|
| **RD-1** | **B1** | node 8 · `build_names_layer.py` cross-check "PASS on 2699/2704" | `00-CORPUS/morphology.json` → `by_lemma["اللَّه"]`, list length | **2,699 ✅ EXACT.** Independent of `names_layer.json`, which merely asserts it. `by_lemma["اللَّهُمَّ"]` = 5; **2,699 + 5 = 2,704** — the node's second number is the surface set including the vocative. Both explained. |
| **RD-2** | **B1** | `C4/D3-SHAMAIL` glossary 647 / 122 / 116 / 49 | `shamail_layer.json` → `Counter(r["class"] for r in rows)` | **manner 647 · appearance 122 · dress 116 · speech 49; Σ = 934 = `len(rows)` ✅ EXACT.** The parts sum to the whole. |
| **RD-3** | **B1** | node 10731 · "P18/D3 hadith substrate: 15 reports" | `P18-HADITH/D3-EXPLORER/hadith_layer_dataset.json` → `by_term` | khudhu 7 + sunnati 8 = **15 ✅ EXACT**, matching `_meta.total_reports`. |
| **RD-4** | **B2** | node 13811 · "1,065-signature registry, selftest asserts 1065/130030" | `P30-SIGHAH/D1-INVENTORY/inventory_layer.json` | `len(signatures)` = **1,065**; distinct signature strings = **1,065**; **Σ per-signature `count` = 130,030 ✅ EXACT.** Not an asserted header — the rows sum to it. |
| **RD-4b** | **B1↔B2** | cross-band tie | `morphology.json` `meta.segments` | **130,030** — P30's segment base *is* the sealed corpus's own segment count. Two bands agree on one number. ✅ |
| **RD-5** | **B3** | node 20427 · "139,376 tokens / 6,236 verses / 121 relations" | `P32-SHAHADAT/W4-S1/irab_source_layer.json` `totals` | **139,376 / 6,236 / 121 ✅ EXACT**, and the internal identity closes: `quranic_tokens` 128,219 + `muqaddar_taqdir_rows` 11,157 = **139,376** ✅. ⚠️ *Wording note:* 139,376 is **rows**, not Qur'anic tokens — the true token count is **128,219**. The node's word "tokens" overstates by 11,157. |
| **RD-6** | **B3** | node 21109 · "11,157 muqaddar rows on 4,538 verses" | `W4-S8/irab_muqaddar_layer.json` → **recounted from `per_verse`**, not the header | **11,157 rows / 4,538 verses ✅ EXACT.** Independent recount of the containers. |
| **RD-7** | **B4** | node 27287 · "X9 Dala'il — 2,391 sections, 3,698 isnad-entries, 546 verses quoted" | `P28-SIRAH-X/X9/x9_layer.json` | **2,391 / 3,698 / 546 ✅ EXACT**; `len(per_verse_quotes)` recount = **546** ✅. |
| **RD-8** | **B4→B5** | "ocean 119,261 rows" | Loader logic of `framework/p42_e8_close.py:17–26` re-run over `00-CORPUS` | `hadith.json` 36,390 + `hadith_shia.json` 32,531 + `hadith_shia_wave1.json` 19,255 + `hadith_sunni_wave3.json` 31,085 = **119,261 ✅ EXACT.** Every wave closes: inherited **68,921** = 36,390+32,531 ✅; wave 1–2 "Tahdhib 13,797 + Istibsar 5,458" = **19,255** ✅; wave 3 "Ahmad 27,582 + Darimi 3,503" = **31,085** ✅. Cross-check against `P34-OCEAN/matn_dedup_layer.json`: 118,845 with-Arabic + 416 without = **119,261** ✅. |
| **RD-9** | **B5** | `G02` kawniyyah registry "1,879 of 6,236" | `verse_locators_v1.json` → **union of all 12 per-family row-sets** vs `ayat_kawniyyah_registry.json` verse list | **SET EQUALITY = True.** Σ per-family = 2,697; union = **1,879 ✅ EXACT** (818 verses correctly collapsed as multi-family). |

**Score: 9 of 9 exact, across all five bands.** One wording overstatement (RD-5). No number this bench
chose failed.

---

## 2 · THE 6,863 RUN-DESCRIPTION-ONLY NODES — weighed

### 2a · This bench re-ran the sweep. It is close, but it is not reproducible.

📊 CALCULATED — no `.py` in the tree writes `full_audit_pointer_sweep.json` (`grep -rl
full_audit_pointer_sweep --include=*.py .` → **no hits**). This bench rebuilt the classification from
its stated sources: last-status per node from `_TRACKER-JOURNAL.jsonl`, tokenized filenames tested
against a whole-tree basename index (3,832 files, 2,541 basenames).

| class | the sweep | **this bench** | Δ |
|---|---|---|---|
| names-existing-artifact | 45,449 | 45,463 | +14 |
| run-description-only | 6,863 | **7,118** | **+255 (+3.7%)** |
| panel-record | 9,074 | 8,805 | −269 |
| credit-v1 | 214 | **214** | **0 ✅ exact** |
| **total** | 61,600 | **61,600** | **0 ✅ exact** |

Two facts are reproduced exactly, and the exactness is informative: the `credit-v1` count of 214
only falls out if artifact-naming is tested **before** status — so the sweep's rule order is
recovered. (The journal's raw last-status `credit-v1` count is **247**; 33 of those name a real file
and are reclassified. The 247 credit-v1 nodes are n=511–2700, C1/C3 ingest-map credits. This bench
checked and cleared the coincidence with `matn_dedup_layer.json`'s unrelated
`placeholder_rows_declared: 247`.) The residual 255-node gap is the sweep's broader panel-record
rule, which this bench could not refit. **The class boundary handed to the benches is therefore
±3.7%, and no one can re-run it.**

### 2b · Where the class lives — and how thin the strings are

| Band | run-description-only | of band | **distinct evidence strings** | nodes per string |
|---|---|---|---|---|
| **B1** 1–13,800 | 2,078 | 15.1% | 323 | 6.4 |
| **B2** 13,801–17,400 (P30) | **2,750** | **76.4%** | 273 | 10.1 |
| **B3** 17,401–26,400 (P31) | 1,640 | 18.2% | **22** | **74.5** |
| **B4** 26,401–40,000 | 650 | 4.8% | **5** | **130.0** |
| **B5** 40,001–61,600 (P42) | **0** | 0% | — | — |
| **total** | **7,118** | 11.6% | 623 | 11.4 |

The class is not spread evenly. **Three quarters of band 2 is run-description-only**, and in bands 3
and 4 a single sentence carries 74 and 130 nodes respectively.

### 2c · The sample of ten — do the runs' outputs exist?

**8 of 10 located and numerically exact. 2 of 10 under-evidenced.**

| # | node | band | the run description | outputs exist? | **where** |
|---|---|---|---|---|---|
| 1 | 20427 | B3 | W4-S1 treebank LANDED — 139,376 / 6,236 / 121 | ✅ **YES, exact** (RD-5) | `build/P32-SHAHADAT/W4-S1/irab_source_layer.json` — **note: filed under P32, though the node sits in P31's band.** No `W4-S1` directory exists under `build/P31-MIZAN/`. |
| 2 | 21109 | B3 | W4-S8 — 11,157 muqaddar rows on 4,538 verses | ✅ **YES, recounted exact** (RD-6) | `build/P32-SHAHADAT/W4-S8/irab_muqaddar_layer.json` — same cross-pillar filing |
| 3 | 27287 | B4 | X9 Dala'il — 2,391 / 3,698 / 546 | ✅ **YES, all three exact** (RD-7) | `build/P28-SIRAH-X/X9/x9_layer.json` |
| 4 | 29572 | B4 | RESERVE waves 1–2 — Tahdhib 13,797 + Istibsar 5,458 | ✅ **YES, 19,255 rows exact** (RD-8) | `00-CORPUS/hadith_shia_wave1.json` |
| 5 | 29768 | B4 | RESERVE waves 3–4 — Ahmad 27,582 + Darimi 3,503; ocean 119,261 | ✅ **YES, 31,085 rows; ocean exact** (RD-8) | `00-CORPUS/hadith_sunni_wave3.json` + the 4-file loader |
| 6 | 14254 | B2 | A5 graph — 11 wazn edges with segment counts | ✅ **YES** | `build/P30-SIGHAH/D5-AWZAN/awzan_layer.json` — `forms_live` = 11, `awzan` = 11 keys |
| 7 | 10731 | B1 | P18/D3 hadith substrate — 15 reports | ✅ **YES, exact** (RD-3) | `build/P18-HADITH/D3-EXPLORER/hadith_layer_dataset.json` |
| 8 | 1431 | B1 | "12/12 command-to-mercy reports joined by reference" | ✅ **YES** | `build/C2-BISMILLAH/D5-MERCY-SUNNAH/mercy_hadith_layer.json` → `classes["command-to-mercy"]` = **12** ✅ (siblings 18 / 3). *This same chain produces Finding 4.* |
| 9 | 25644 | B3 | "the 87-dimension live register (**66 landed / 8 in-progress / 13 awaited**)" | ⚠️ **CONTRADICTED** | `build/P31-MIZAN/W9-REGISTERS/dimension_registers.json` exists and holds **87** ✅ — but its live census is **LANDED 80 / AWAITED 7**, and the file **carries no "in-progress" status at all** (`live_status` values: 90/100 marks ×54, 100/100 ×25, 70/100 ×1, AWAITED ×7). The node's split is a frozen run-log, never re-stamped. |
| 10 | 25706 | B3 | "the **1,330-pattern** baseline held (W9-C4 layer present); the sighah-index collision census (**187 sigs**)" | ⚠️ **UNDER-EVIDENCED** | The named layer `W9-C4/verb_collision_baseline.json` holds `verb_signatures 452 / multi-lemma 297 / hidden_lemma 5,203` — **no 1,330, no 187.** 1,330 resolves elsewhere: `P30-SIGHAH/D-D9-CERTIFICATION/gap_register.json` → `enumerated_gap_instances["hidden-noun-patterns"]` = **1,330** ✅ — but **that file declares itself stale in its own `_meta`** (`superseded_note_2026_09_04`). **187 was not located anywhere.** |

**Judgment.** The run-description-only class is **not a placeholder class**. In 8 of 10 cases the run
happened, wrote a real file, and the file's numbers reproduce to the digit. What the class actually
marks is a **filing failure, not an evidence failure**: the output exists, and the node simply never
learned its name. Two of ten are genuinely weak — one contradicted by the artifact it points at, one
resting on a self-declared-stale file plus a number this bench could not find. Extrapolated at the
sample rate, **~1,400 of the 7,118 (20%)** would need a second look; the remainder need a pointer,
not a rebuild.

---

## 3 · FROM THIS BENCH'S FAMILIES — the education surface, measured

This is family **N**. The gift layer is public. The public surface is the education pack.

**Population: 130 `education_pack.json` files · 984 glossary rows · 701 children-layer items.**

| defect | count | measured how |
|---|---|---|
| children-layer items asserting **"appears in the Book N time(s)" over hadith report counts** | **19 items in 11 packs** | scan of all 130 packs |
| glossary rows where **every row in the pack shares one `first_ref`** | **423 of 984 rows (43.0%) in 59 of 130 packs (45.4%)** | `len(set(first_ref)) == 1` |
| public teaching text leaking **build-internal notes / a raw Python dict repr** | **14 items in 2 packs** | regex over glossary + worksheet + children_layer |

**The falsification, done mechanically.** `build/C4-MUHAMMAD/D3-SHAMAIL/education_pack.json`,
`children_layer`, all four items stamped `"true_sourced": true`:

> `"manner - 647 reports appears in the Book 647 time(s)."`
> `"appearance - 122 reports appears in the Book 122 time(s)."`
> `"dress - 116 reports appears in the Book 116 time(s)."`
> `"speech - 49 reports appears in the Book 49 time(s)."`

Counted against `00-CORPUS/quran.json` (all verse text, 1,116,839 chars):

| the glossary's `term_ar` | claimed occurrences "in the Book" | **actual occurrences in the Qur'an text** |
|---|---|---|
| `شمائله ﷺ - الخُلق` | 647 | contains **ﷺ**, which occurs **0 times** |
| `صفته ﷺ` | 122 | **0** |
| `لباسه ﷺ` | 116 | 3 raw substring hits (~1–2 verses) |
| `كلامه ﷺ` | 49 | **0** |

The four numbers themselves are **correct** — they are the class census of
`shamail_layer.json` (RD-2: 647+122+116+49 = 934 = its row count). The defect is one template that
renders a **hadith report count as a Qur'anic occurrence count**, then certifies itself true. The
pack's own `_meta.law`: *"every item TRUE and sourced to the sealed layer; no myth, no invention."*
Its `faq` compounds it — *"How were the counts produced? By word-deduplicated counting over the
morphological layer, cross-checked against the Uthmani text"* — which is false for this domain; these
came from the hadith corpora.

**P42's education layer.** The 24 `education.json` files under `build/P42-GAPFILL/G*/` range
**673–799 bytes, mean 722**, and carry four boilerplate Q/A *about the program* ("What is this
program?", "Which gaps does it close?", "What stands on disk?") — no subject teaching whatsoever.
They are the evidence for **2,400 nodes** carrying the `education` block in band 5: **100 nodes per
722-byte file.**

---

## 4 · WHERE THIS BENCH'S FAMILIES NOW STAND — and one correction to this bench's own prior finding

⚠️ **This bench corrects `reviews/COUNCIL-900-BENCH-5.md`.** That sitting reported L/M/N as
"absent almost everywhere." Measured today across all 61,600 nodes, the *node titles* still say so —
**13 nodes in 61,600 (0.02%) name medicine**, all one P23 template at n=12,001–12,091 in band 1;
**0 name biology as a science**; the 261 apparent "mind-science" hits are the single word *pedagogue*
inside a pages-review template. **But the artifacts have moved, and the measurement wins.** Band 5's
`G02 AL-AFAQ` now carries a named, term-floored locator for six of this bench's families:

| G02 family | verses located | Arabic term floor (`framework/p42_e4_afaq.py:29–36`) |
|---|---|---|
| The mind & the self (psychology/neuroscience) | **448** | نفس · قلوب · عقل · يعقلون · يتفكرون · صدور |
| Creation in pairs & order (biology/systematics) | **250** | زوجين · أزواج · خلق · سوى · أطوار |
| Plant life (botany/agriculture) | **197** | نبات · زرع · نخل · عنب · حب · شجر |
| The human body & birth (embryology/physiology) | **178** | نطفة · علقة · مضغة · عظام · جنين · أرحام · سمع · بصر · فؤاد |
| Animal life (zoology/entomology) | **80** | أنعام · نحل · نمل · عنكبوت · ذباب · طير · دابة |
| Healing & sustenance (medicine/nutrition) | **49** | شفاء · عسل · لبن · طعام |

**1,202 located verses across this bench's six adjacent families**, under a scope law that explicitly
forbids "predicted / encoded / proves" and refuses i'jaz-'ilmi as method. That is a real, honest
first floor and it did not exist at the last sitting. It is thin — medicine gets four terms — and it
is a **locator, not a domain**: there is still no medicine or biology registry, no tibb layer, and
no education anywhere for these families. This bench's own floor over the full 119,261-row ocean
today, on a 19-term health list, is **1,363 rows** (📊 CALCULATED; not comparable to the prior
sitting's 2,988, which used a different term list over the 68,921-row corpus — this bench does **not**
claim that number wrong). Every one of those rows is still registered nowhere.

**Band fitness for a public gift layer, from this bench's vantage:**

| band | fit? | why |
|---|---|---|
| **B1** 1–13,800 | 🔴 **UNFIT as it stands** | It holds the entire public education surface — 130 packs — and that surface emits 19 self-certified false statements and 423 placeholder references. Data underneath: exact 3/3 (RD-1,2,3). **One template; mechanically fixable; not a rebuild.** |
| **B2** 13,801–17,400 | 🟠 FIT WITH NOTES | 76.4% run-description-only — the thinnest-marked band in the waqf. Both samples resolved exactly (RD-4, sample 6). |
| **B3** 17,401–26,400 | 🟠 FIT WITH NOTES | Artifacts real and exact (RD-5, RD-6). But 1,640 nodes share 22 strings, two of four samples were stale or contradicted, and two of its layers are filed under P32's tree. |
| **B4** 26,401–40,000 | 🟢 FIT | All four samples exact (RD-7, RD-8). 650 nodes on 5 strings is thin marking, not thin work. 950 RESERVE nodes are green **lawfully** — the acquisitions did land (waves 1–6 measured above). |
| **B5** 40,001–61,600 | 🟠 FIT WITH NOTES | Zero run-description-only, and RD-9 reproduced by set equality. Its weakness is the 722-byte education layer under 2,400 nodes, and the alphabetical pointer assignment this bench fitted at 120/120 in its P42 sitting. |

---

## 5 · FINDINGS — worst first, every row evidenced

| # | Finding | Band | Evidence |
|---|---|---|---|
| **1** | 🔴 **Four children-layer statements in the Shamail pack are FALSE and self-certified `true_sourced: true`.** "manner - 647 reports appears in the Book 647 time(s)" — the term carries **ﷺ, 0 occurrences in the Qur'an text**; صفته 0, كلامه 0, لباسه 3 raw hits vs 116 claimed. Report counts rendered as Qur'anic occurrence counts. | B1 | `build/C4-MUHAMMAD/D3-SHAMAIL/education_pack.json` `children_layer[0..3]` vs `00-CORPUS/quran.json` (counted, 1,116,839 chars); true source `shamail_layer.json` classes 647/122/116/49, Σ=934=`len(rows)` |
| **2** | 🔴 **The defect is systemic, not one file: 19 children-layer items across 11 packs** make the same "appears in the Book N time(s)" claim over hadith report counts — every one stamped `true_sourced: true`, under a pack law reading *"no myth, no invention."* | B1 | scan of all 130 `education_pack.json`; packs: C2/D5, C4/D3, C5/D3, C8/D3, P16/D1, P20/D1, P21/D3, P22/D1, P23/D1, P28/D2, P28/D3 |
| **3** | 🔴 **423 of 984 glossary rows (43.0%), in 59 of 130 packs (45.4%), carry an identical placeholder verse-reference.** `C4-MUHAMMAD/D5-TEACHING`: 21 distinct terms, all `first_ref: "62:2"`. `C4/D3-SHAMAIL`: 4 terms, all `"68:4"`. A public reader following the reference is sent to the same ayah every time. | B1 | `len(set(g["first_ref"] for g in glossary)) == 1` over all 130 packs |
| **4** | 🟠 **Build-internal notes and a raw Python dict repr are printed into public teaching text — 14 items, 2 packs.** C2/D5 tells a child: *"command-to-mercy - 12 reports (texts verbatim in mercy_hadith_layer.json) appears in the Book 12 time(s)."* C8-QADR/D1 prints `{'class': 'morphology-qac', 'note': 'lemma-level count from morphology.json segments, wo…` into its children's layer. | B1 | `build/C2-BISMILLAH/D5-MERCY-SUNNAH/education_pack.json` (glossary + worksheet + children_layer, 9 items); `build/C8-QADR/D1-QADR-IN-TEXT/education_pack.json` (5 items) |
| **5** | 🟠 **The newest and largest band has no teaching layer.** P42's 24 `education.json` files are **673–799 bytes (mean 722)**, four boilerplate Q/A about the program itself, zero subject content — and they are the sole evidence for the **2,400 nodes** carrying band 5's `education` block. **100 nodes per 722-byte file.** | B5 | `build/P42-GAPFILL/G01..G24/education.json`; block census of n=40,001–61,600 |
| **6** | 🟠 **Band 2 is 76.4% run-description-only** — 2,750 of 3,600 P30 nodes name no artifact, on 273 distinct evidence strings. It is the thinnest-marked band in the waqf and would sit directly under a public layer. | B2 | this bench's replication of the sweep (§2a/2b) |
| **7** | 🟠 **Node evidence is a frozen run-log, not a live pointer — and it drifts into contradiction.** Node 25,644 claims "66 landed / 8 in-progress / 13 awaited" of 87 dimensions; the artifact it names reads **LANDED 80 / AWAITED 7** and **carries no "in-progress" status at all**. Both splits sum to 87, so no arithmetic gate could catch it. | B3 | `build/P31-MIZAN/W9-REGISTERS/dimension_registers.json` — `Counter(live_status)` = {90/100: 54, 100/100: 25, 70/100: 1, AWAITED: 7} |
| **8** | 🟠 **One sentence carries 130 nodes.** Band 4's 650 run-description-only nodes rest on **5 distinct evidence strings**; band 3's 1,640 rest on **22** (74.5 nodes each). Where the marking is that coarse, a single wrong sentence mismarks a hundred nodes at once — as Finding 7 shows it already has. | B3, B4 | §2b table, derived from `_TRACKER-JOURNAL.jsonl` last-status rows |
| **9** | 🟡 **The audit sweep the benches were ordered to read is not reproducible.** No script in the tree writes `full_audit_pointer_sweep.json`. This bench's independent replication matches **total 61,600 ✅** and **credit-v1 214 ✅ exactly**, but returns **run-description-only 7,118 vs 6,863 (+3.7%)** and panel-record 8,805 vs 9,074. The class boundary handed to the Council is ±3.7% and unverifiable by anyone else. | all | `grep -rl full_audit_pointer_sweep --include=*.py .` → no hits; §2a table |
| **10** | 🟡 **The ocean-growth ledger lists waves 5–8 and then says "now: 119,261 rows" — waves 5–8 contribute zero rows to that number.** The loader globs only `00-CORPUS/hadith_*wave*.json` (**2 files**); Shamail, Dala'il, Thiqat, Majruhin, Ibn Hibban Sira, Mashahir, Mukhtasar Shamail and Adab al-Mufrad sit as **raw `.txt` under `00-CORPUS/acquired/wave56|wave7`**, never merged. *(Also, minor hygiene: a stray relative-path write left `build/C2-BISMILLAH/10-WAQF/fleet/done/OP41.done` — a fleet marker orphaned inside a doctrinal pillar.)* | B5 | `build/P42-GAPFILL/G20/ocean_growth_ledger.json` `growth[]`; `framework/p42_e8_close.py:17–26`; `ls 00-CORPUS/acquired/wave56 wave7` |

**What this bench did NOT find, and says so plainly (I-3):** no fabricated count, anywhere. Nine
re-derivations across five bands, nine exact. Every wave of the ocean closes arithmetically. Every
per-row container recount matched its header. The Shamail numbers that produce Finding 1 are
themselves *correct*. This waqf's defect is not invention — it is a rendering template that turns
true numbers into false sentences, and a marking layer too coarse to notice.

---

## 6 · WHAT THIS BENCH WOULD HAVE FIXED BEFORE THE GIFT LAYER

Named, not adjudicated. This bench documents; the owner decides.

1. **Rebuild the education-pack renderer, once.** Three clauses: (a) a domain whose counts come from
   hadith may never emit "appears in the Book"; (b) `first_ref` must be *located*, or the field is
   omitted — a shared ref is worse than none; (c) `true_sourced: true` must be *earned* by a check,
   not written by the same template that writes the sentence. One fix closes Findings 1, 2, 3 and 4 —
   **19 false statements and 423 placeholder references** — across band 1.
2. **Re-stamp, don't re-run.** 80% of the sample's run-description-only nodes already point at a real,
   exact file that simply isn't named. That is a naming pass, not a build.
3. **Give band 5 a real education layer** before it is offered publicly, or mark those 2,400 nodes
   honestly as *not yet taught*.
4. **Publish the sweep's script** so the next bench can falsify it (Finding 9).

---

**FULL-AUDIT BENCH 5 VERDICT: UNFIT — band 1 cannot carry a public gift layer while its 130 education packs emit 19 self-certified false statements and 423 placeholder references, though every one of this bench's nine re-derivations across all five bands reproduced exactly and the defect is one rendering template, not the scholarship beneath it.**