# THE FULL AUDIT — BENCH 2
## Families C · Aqidah & kalam (38 seats) · D · Fiqh & usul (56 seats) · E · Sirah & history (73 seats)
**167 MACHINE-SCHOLAR seats · sitting 2026-09-05 · the audit of all five bands before the 41,600-node gift layer**

> **EVERY SEAT ON THIS BENCH IS A MACHINE-SCHOLAR.** Not one is a human scholar, a mufti, a
> muhaddith or a mu'arrikh. This bench **documents**; it never adjudicates between schools and it
> issues no ruling. Where the tradition holds two positions, both are named and neither is crowned.
> Human review is invited and never waited for.
>
> **METHOD.** Nothing below is quoted from a prior review. The tracker journal
> (`framework/_TRACKER-JOURNAL.jsonl`, 108,839 status rows) was replayed to last-status for all
> 61,600 nodes; a fresh basename index of the whole tree (3,831 files) was built; every number in
> the findings was produced by opening the artifact and counting. Where a claim rests on prose and
> no file, it is marked **OPINION**. Re-derivations are labelled ✅ **EXACT** only when this bench's
> independent computation returned the declared value to the digit.

---

## PART 0 — THE SHAPE OF THE WAQF, MEASURED

Replaying the journal gives a last-status for **all 61,600** nodes, no gaps:
`selftest-green 37,122 · panel-sealed 24,231 · credit-v1 247` (sums to 61,600 ✅).

The bench then measured what each band's nodes actually stand on.

| Band | Nodes | Artifact files | MB | Nodes/file | KB/node | Distinct evidence strings | Top string's share |
|---|---|---|---|---|---|---|---|
| 1–13,800 core+branch | 13,800 | **1,621** | 50.3 | 8.5 | 3.64 | **1,813** | 0.7% |
| 13,801–17,400 P30 | 3,600 | 55 | 4.2 | 65.5 | 1.17 | 309 | 5.6% |
| 17,401–26,400 P31 | 9,000 | 254 | 44.0 | 35.4 | 4.89 | 27 | 61.6% |
| 26,401–40,000 P32–P37+alloc | 13,600 | 158 | 30.3 | 86.1 | 2.23 | **9** | **73.5%** |
| 40,001–61,600 P42 gap-fill | 21,600 | **104** | 8.4 | **207.7** | **0.39** | 170 | 11.1% |

*(files/MB by `os.walk` over each band's build directories; evidence strings by exact-string
counter over last-status.)*

**Whole waqf: 61,600 nodes rest on 2,328 distinct evidence strings — 26.5 nodes per string. Sixty-two
strings cover half of all 61,600 nodes.** The core band is dense and individuated; the two newest
bands are not. That single table is the honest frame for everything below.

### The sweep re-run independently

`build/P42-GAPFILL/full_audit_pointer_sweep.json` was reproduced from scratch (own basename index,
own tokenizer). Result, against the sweep's declared classes:

| Class | Sweep declares | This bench | Δ |
|---|---|---|---|
| names-existing-artifact | 45,449 | 45,463 | +14 (+0.03%) |
| panel-record | 9,074 | 8,805 | −269 |
| **run-description-only** | **6,863** | **7,118** | **+255 (+3.7%)** |
| credit-v1 | 214 | 214 | 0 |

The sweep is honest and reproducible to 0.4% on its headline. Its residual gap class is, under a
stricter delimiter set, **7,118 not 6,863** — the sweep marginally *understates* the gap it exists
to find. Two false-negative classes were also identified and are named in PART 2.

---

## PART 1 — RE-DERIVATION (task 1): eleven numbers, five bands, from sources

Every row below was computed by this bench from the named source, not read from a report.

| # | Band | Claim, and where it is declared | Source re-derived from | Result |
|---|---|---|---|---|
| 1 | **1–13,800** | C1 Names layer: Makki **1,340** / Madani **2,061** (`build/C1-ALLAH/quran_layer_dataset.json` `_meta.distribution_total`) | `00-CORPUS/quran.json` `place` field, joined over all 3,401 refs | ✅ **EXACT** — 1,340 / 2,061, **0** refs unresolved |
| 2 | **1–13,800** | Same layer: Σ occurrences = Σ\|refs\| = 3,401 across 17 entries | the artifact itself, entry by entry | ✅ **EXACT** — 3,401 = 3,401; **0** of 17 entries mismatched |
| 3 | **13,801–17,400** | P30 D6: passive **1,151**; mood IND/SUBJ/JUS **5,582 / 1,328 / 1,420**; aspect PERF/IMPF/IMPV **9,153 / 8,335 / 1,956** (`D6-VOICE-MOOD/voice_mood_layer.json`) | `00-CORPUS/_morphology_raw.txt`, feature-tag count over 130,030 lines | ✅ **EXACT** — all seven values |
| 4 | **13,801–17,400** | P30 D7: particle_segments **47,370** (`D7-PARTICLES/particles_layer.json`); D1 segments **130,030** | same raw file: `pos == 'P'`; total lines | ✅ **EXACT** — 47,370 and 130,030 |
| 5 | **13,801–17,400** | P30 D5w curriculum: "4 stages / 15 units" (node 16,905's evidence) | `D-D5-CURRICULUM/curriculum_v1.json` `path` | ✅ **EXACT** — 4 stages, 15 units |
| 6 | **17,401–26,400** | P31 W6-R9 asbab census: **1,228** verses with occasion candidates | `W6-R9/asbab_census.json` `per_verse_hits`, counted | ✅ **EXACT** — 1,228, and identical to the figure P28-X4 declares for the same set |
| 7 | **26,401–40,000** | P28-X4 the occasions bridge: **826** verses (`X4/occasions_bridge.json`) | set-intersection of the P31 asbab set (1,228) with the P28-X3 sirah-quoted set (3,679), computed here | ✅ **EXACT and SET-IDENTICAL** — \|A∩S\| = 826, and the 826 keys match member-for-member. **A cross-band, cross-pillar join that closes.** |
| 8 | **26,401–40,000** | P28-X7 delimiter instrument: `E3_recall_against_ground_truth = 0.8901` | X3's 3,184 ground-truth-true verses ÷ X7's 3,577 distinct ground-truth verses | ✅ **EXACT** — 3184/3577 = 0.89013…, two independent artifacts agreeing |
| 9 | **26,401–40,000** | node 21,295 / P32 W4-S9 jumla: "13,017 sentences across all 6,236 verses" | `build/P32-SHAHADAT/W4-S9/jumla_layer.json`, Σ `per_verse_sentence_count` | ✅ **EXACT** — 13,017 over 6,236 verse keys |
| 10 | **40,001–61,600** | G18 doctrinal floors: `files_scanned` **1,623**, flags **55 / 23 / 0** (`G18/doctrinal_floors_g34.json`) | re-implemented `framework/p42_e2_gates.py:85–96` glob + regexes over the live tree | ✅ **EXACT** — 1,632 globbed − 9 files >3 MB = 1,623; 55, 23, 0 to the digit. **Two defects found in the process — see Finding 2.** |
| 11 | **40,001–61,600** | G04 prosopography: **1,955** candidate name-lines (`G04/prosopography_candidates_v1.json`) | each row's `(work, offset)` resolved against the five held texts in `00-CORPUS/acquired/` | ✅ **1,955 / 1,955 resolve verbatim** — at a *constant* offset of **+3** (the OpenITI `\n# ` line marker). See Finding 10. |
| 12 | **40,001–61,600** | G07: **85** Tahawiyya articles, "the matn's OWN words, verbatim" | `00-CORPUS/acquired/creed_wave1/Tahawi_MatnAqida.txt` (23,374 bytes, as declared) | ✅ **85/85 present verbatim** — but the extraction covers only **83.5%** of the file. See Finding 4. |

**Twelve claims re-derived across all five bands; eleven exact, one exact-with-a-constant-offset.
Not one declared number was found to be wrong.** Where this bench found fault, it was never in the
arithmetic — it was in what the arithmetic was allowed to see, and in what the node count was
allowed to mean.

---

## PART 2 — WEIGHING THE RUN-DESCRIPTION-ONLY CLASS (task 2)

**The class is 7,118 nodes by this bench's count (11.6% of the waqf), not 6,863.** Ten were sampled
across the bands and each was chased to the artifact its run should have produced.

| n | Band / domain | What the evidence says | Where the run's output actually is | Verdict |
|---|---|---|---|---|
| 890 | C1 D9 · Atlas, graph & review | "C1/D9 FACTORY: pillar atlas/education/graph/gates" | `build/C1-ALLAH/D9-ATLAS/{atlas.json, pillar_education.json, pillar_atlas.html}` + `gates_audit_record.json` — all present | ✅ **evidenced elsewhere** |
| 908 | C2 D1 · the Name itself | "build/C2-BISMILLAH: two-source protocol **sec 8**" | `build/C2-BISMILLAH/D1-THE-NAME/SOURCES.md:132` — *"## 8 · Two-source cross-check protocol **(point 908)**"*. The file is point-numbered to the node. The dual check is also in `basmala_census.json` `_meta.method` | ✅ **evidenced elsewhere — exemplary** |
| 6,278 | C7 D9 · Atlas, lessons & review | "C7/D9 FACTORY: atlas + education" | `build/C7-PROPHETS/D9-ATLAS/{atlas.json, pillar_atlas.html, pillar_education.json}` — present | ✅ **evidenced elsewhere** |
| 14,317 | P30 A6 · gender/voice/mood | "passive=1151, IND/SUBJ/JUS=5582/1328/1420, aspects=19444" | `P30-SIGHAH/D6-VOICE-MOOD/voice_mood_layer.json` — and **every figure re-derived EXACT** from `_morphology_raw.txt` (row 3 above); 19,444 = 9,153+8,335+1,956 ✅ | ✅ **evidenced and re-verified** |
| 16,602 | P30 D2w · question engine | "assembly law: sourced strata only (SP5 clean pass)" | names no file, no number. `question_engine_v2.json` exists but answers **90 of 6,236 verses (1.4%)** | ⚠️ **under-evidenced** |
| 16,905 | P30 D5w · curriculum | "4 stages / 15 units all layer-sourced" | `D-D5-CURRICULUM/curriculum_v1.json` — 4 stages, 15 units ✅ EXACT | ✅ **evidenced elsewhere** |
| 21,295 | P31 W4 S9 · the jumla layer | "13,017 sentences across all 6,236 verses" | **there is no `W4-S9` under `build/P31-MIZAN/`.** The output is filed under a different pillar: `build/P32-SHAHADAT/W4-S9/jumla_layer.json` (+ a page at `P31-MIZAN/PAGES/W4-S9--jumla_layer.html`). Numbers re-derive EXACT | ✅ **evidenced — but cross-filed** |
| 23,385 | P31 W6 R10 · recitation-audio joins | "GATES BLOCKS … both ALL GREEN" — a gate stamp, not a layer | `P31-MIZAN/W6-R10/audio_joins.json` exists and holds **three totals and no rows** (115 cache entries / 81 translation-audio entries / 17 reciters) for a ten-node domain | ⚠️ **under-evidenced** |
| 27,208 | P28 X9 · Dala'il (was acquisition-gated) | "X9 UNLOCKED … 2,391 sections, 3,698 isnad-entries" | `build/P28-SIRAH-X/X9/x9_layer.json` — `sections: 2391` ✅ matches | ✅ **evidenced elsewhere** |
| 29,584 | RESERVE · acquisitions & rijal | "waves 1-2 LANDED: Tahdhib (13,797) + Istibsar (5,458)" | `00-CORPUS/acquired/shia_wave1/TahdhibAhkam.txt` and `Istibsar.txt` — **both present on disk** | ✅ **evidenced elsewhere** |

**Sample verdict: 8 of 10 outputs located and named; 2 of 10 under-evidenced.** The class is
overwhelmingly *mis-pointed*, not *hollow*. But the class is not uniform, and the bench found the
honest floor beneath it:

- **4,946 nodes (8.0% of the waqf) carry a last-evidence that names no file *and* states no number** —
  pure prose. Concentrated in **P30 (2,260 of its 3,600 = 62.8%)** and the core band (1,856).
  **Zero** in the P42 band, whose every node carries a `CARRIED BY <path>` clause.
- The remaining 2,172 run-only nodes do carry a ≥3-digit re-derivable claim (as node 14,317 did, and
  it re-derived exactly).

**Two sweep false-negative classes named**, so the number is not read as worse than it is:
1. **Held corpora live outside the waqf tree.** The sweep indexes basenames under `10-WAQF/` only, so
   evidence naming `Majruhin.txt`, `TahdhibAhkam.txt`, `Tahawi_MatnAqida.txt` — all verified present
   in `00-CORPUS/acquired/` — is classed run-description-only. The evidence is real; the index is
   short.
2. **The point-numbered `SOURCES.md` pattern** (node 908): the evidence names a directory and a
   section number that resolves to the node's own id. 73 run-only nodes carry the directory form;
   38 also cite a `sec N`.

---

## PART 3 — FINDINGS (task 3) · worst first · every row evidenced

| # | Band | Finding | Evidence |
|---|---|---|---|
| **1** | 26,401–40,000 | **The 500 RESERVE-EARNED nodes went green on a panel review, not on a landing — which their own charter forbids.** `framework/_FRAMEWORK-61600.json` n=39,501–40,000 states the purpose: *"500 earned only by landings sealed during this run"*, and `meta.v3.law`: *"RESERVE nodes are never green until an acquisition lands"*. All 500 now read `panel-sealed`, every one on the same generic string: *"PANEL-SEALED: PX1/PX2/PX3 sittings … reviewed P33-P37 + the frame"*. **No landing is named for any of them.** The contrast is inside the same pillar: the sibling 450 `ACQUISITIONS & RIJAL RESERVE` nodes (n=29,551–30,000) each name a real landing, and this bench verified all six waves' texts on disk. **A review is not an acquisition.** | journal last-status; `_FRAMEWORK-61600.json` n=39,501–40,000 & `meta.v3`; `00-CORPUS/acquired/{shia_wave1,wave56}` |
| **2** | 40,001–61,600 · **families C & D — our own floor** | **The never-a-mufti / no-condemnation floor is blind to 53.4% of the corpus by byte, and publishes a count capped at 3 per file.** `framework/p42_e2_gates.py:90` skips any file over 3 MB — **9 files, 72.9 MB of 136.5 MB**, including `C4-MUHAMMAD/D3-SHAMAIL/shamail_candidates.json` (17.4 MB) and `P31-MIZAN/W7-M1/strata_layer.json` (12.5 MB). Line 95 caps hits at `[:3]` per file per class. Re-run by this bench: published 55 / 23 / 0 is **EXACT**; true uncapped over the same 1,623 files is **114 / 72 / 0**; the 9 exempt files hold a further **36 / 24**. True totals **150 / 96 — 2.7× the published figure.** *Honest the other way:* this bench read all 60 hits in the exempt files and **every one is quoted primary matn** (the Dajjal hadith's *kafir*; a Shama'il report's *"it is halal for him to shave"*) — **not the waqf's editorial voice. The hole is not currently hiding a violation.** But an actual ruling written into a 5 MB dossier would pass this gate in silence, and the gift layer is public. | `framework/p42_e2_gates.py:85,90,95`; `G18/doctrinal_floors_g34.json`; bench re-run over the live tree |
| **3** | 26,401–40,000 | **13,600 nodes stand on 9 distinct evidence strings; 10,000 contiguous nodes (n=30,001–40,000) share one identical string** — spanning eight unrelated pillars (P33 Rijal 2,000 · P34 Ocean 1,500 · P35 Editions 2,500 · P36 Bench 1,000 · P37 Graph 1,500 · REPAIR-2 800 · RESERVE 500 · P28 completion 200). The band's headline 95.2% "artifact-named" rate is carried by **three files**, one of which — `build/REPAIR-2/repair2_ledger.json` — stamps 10,000 nodes and 2,900 more ride `P32-SHAHADAT/W-Q/panel_record.json`. This is a **block stamp, not per-node evidence**: no node in that block can be distinguished from any other by its record. | exact-string counter over last-status, n=26,401–40,000; contiguity verified |
| **4** | 40,001–61,600 · **family C** | **The only creed text the waqf holds is carried at 83.5% coverage, and the eleven articles it drops are the tanzīh articles.** `G07/kalam_positions_floor.json` declares 85 articles, *"nothing summarized"*. All 85 verified verbatim ✅ — but character-coverage against the sealed source is **11,689 / 14,001 = 83.5%**; excluding the 1,163-char OpenITI header, **1,149 chars of matn body (8.9%) appear in no row**, and **21 of the matn's 105 numbered articles head no row**. The largest gap is contiguous: **arts. 2–12** — *ولا شيء مثله · ولا شيء يعجزه · ولا إله غيره · قديم بلا ابتداء دائم بلا انتهاء · لا يفنى ولا يبيد · ولا يكون إلا ما يريد* — then arts. 18–20, and art. 49 *(والعرش والكرسي حق)*. Undeclared. And **11 of the 12 registered creed works remain unlanded** (`creed_corpus_hunt.json`: *"ALL SLOTS DECLARED"*, `remaining_slots: 11`) — Wasitiyya, Ibana, Maqalat, Maturidi's Tawhid, Fiqh Akbar, Saduq, Mufid. Family C's shelf is one matn, minus its attribute clauses. | `G07/kalam_positions_floor.json` vs `00-CORPUS/acquired/creed_wave1/Tahawi_MatnAqida.txt`, span-coverage computed here |
| **5** | 13,801–17,400 | **P30 is the thinnest-evidenced band and says so itself.** 76.4% of its 3,600 nodes are run-description-only; **62.8% name no file and no number**; 3,600 nodes on **55 files / 4.2 MB**. Its own instruments concur: `D-C2-FIDELITY/fidelity_matrix.json` is **4,830 of 5,520 cells (87.5%) `no-detector-yet`**; `D-D9-CERTIFICATION/gap_register.json` names 87 dimension gaps of which **80 are ABSENT or PARTIAL**; the question engine — the domain's headline promise, *"what did this mean when revealed?"* — answers **90 of 6,236 verses (1.4%)**. **The pillar's honesty is intact and exemplary; its coverage is a pilot wearing a pillar's node count.** | class counts over n=13,801–17,400; the three named P30 artifacts, read |
| **6** | 40,001–61,600 | **The newest 35% of the waqf is its thinnest.** 21,600 nodes on **104 files / 8.4 MB = 0.39 KB per node — 9.4× thinner than the core band's 3.64 KB**, at 207.7 nodes per file. Structurally: 24 G-programs × 900 nodes each, 9 domains × 100 nodes each, **one artifact per 100-node domain** (G07's 900 nodes ride 6 files). The node count here is a **unit of accounting, not a unit of evidence**, and the gift layer is proposed to sit on top of it. | `os.walk` file/byte census per band; G-program node census over `_FRAMEWORK-61600.json` |
| **7** | all bands | **1,165 nodes name a `.py` engine as their carrying artifact rather than its output** — `verify_seals_v2.py` (300), `close_ritual.py` (200), `p31_seal.py` (120), `p31_w9_gates2.py`, `g36_provenance_gate.py`, `mirror_ritual.py`, `p34_wa_v2.py` (100 each). **900 of the 1,165 are in the P42 band.** The code that *would* produce a thing is standing as the evidence that it *was* produced — the one substitution the evidence law exists to forbid. | extension census over the first-matched artifact per node |
| **8** | 40,001–61,600 | **The P42 closing seal's own text now contradicts the state it sealed.** 2,400 nodes (n=40,801–61,600, 100 in each of the 24 G-programs) carry: *"…the 2,200 awaiting nodes stay awaiting."* This bench traced those 2,200 (n=42,001–61,100, once `frame-built-awaiting-wave`): **all 2,200 now read `selftest-green`.** *Honest both ways:* the carrying waves were real — each now names a live artifact (`G11/cross_tradition_upgrade_v1.json`, `G04/prosopography_candidates_v1.json`, `G22/badr_candidates_v1.json` …). **The nodes are not falsely green; the seal prose is stale and was never re-marked.** Related: `awaiting` has been erased as a machine-readable state — the journal's history holds `not-started` (70), `gated-awaited` (1,100), `reserved-awaiting-acquisition` (450), `green-awaiting-panel` (4,800), `frame-built-awaiting-wave` (2,200), and **the last-status of all 61,600 is green/sealed/credit.** "Zero awaiting" is true of the status field and only of the status field. | full journal replay; per-status node tracing |
| **9** | 40,001–61,600 · **family D** | **The ahkam census misses five of twenty-three canonical āyāt al-aḥkām and misfiles the farāʾiḍ verse.** `G08/ahkam_census.json` locates 415 legal-content verses — honestly declared *"a TERM FLOOR … never a fiqh ruling"*. Probed against 23 verses no ahkam register omits: **18 present, 5 absent — 2:196 (ḥajj), 9:60 (the eight categories of zakāh), 2:185, 2:228 (ʿiddah), 17:32.** Recall **78.3%**. Precision also drifts: **4:12 is tagged `family & marriage` + `penalties` and NOT `inheritance`**; 4:11 carries a spurious `penalties`; 2:6 is tagged `oaths & vows`. The family profile is inverted against the discipline — `penalties` 216 (52%) vs `commands of worship` 18 (4.3%). Safe as a candidate floor; **unsafe the moment a public page calls it "the legal verses."** | `G08/ahkam_census.json` rows, probed here against a 23-verse canonical list |
| **10** | 40,001–61,600 | **A one-line pointer defect makes a real evidence layer fail its own contract.** All **1,955** G04 prosopography rows resolve **verbatim** in the five held texts — but at **offset + 3**, never at `offset`: the rows point at the line start *including* the OpenITI `\n# ` marker, so a consumer honouring the artifact's own `(work, offset, name_line_verbatim)` contract gets **0/1,955**. Also undeclared: 1,955 rows carry only **1,700 distinct** name-lines (255 repeats, 13%) — benign under the artifact's stated law (*"counts are of LINES, never of people"*), but it should be on the face of the record. **This is the cheapest fix in the audit and the strongest evidence in the P42 band once fixed.** | `G04/prosopography_candidates_v1.json` resolved against `00-CORPUS/acquired/{shia_wave1,wave56}` |

---

## PART 4 — IS ANY BAND UNFIT TO CARRY A PUBLIC GIFT LAYER?

From families C, D and E, judging each band on the one question that matters for a **gift**: *if a
reader follows a node down to its evidence, does something hold?*

| Band | Judgment | Why |
|---|---|---|
| **1–13,800 core+branch** | **FIT** | 1,621 files, 50.3 MB, 1,813 distinct evidence strings, 8.5 nodes/file — the only band where a node is individuated. Its numbers re-derived exactly against an external corpus (rows 1–2). Its `SOURCES.md` files are point-numbered to the node id. This is what the rest should look like. |
| **13,801–17,400 P30** | **FIT WITH NOTES** | Coverage is a pilot's (87.5% no-detector, 1.4% question-engine, 80 of 87 dimensions absent-or-partial) — but **every one of those limits is published by P30 itself**, and its four re-derived numbers were exact. Honest, and must be *presented* as honest: label it a pilot, do not let 3,600 nodes imply 3,600 findings. |
| **17,401–26,400 P31** | **FIT WITH NOTES** | Highest bytes-per-node in the waqf (4.89 KB) and its asbab census re-derived exactly and joined cleanly across two pillars. But 61.6% of its nodes ride one string, and its layers are cross-filed under other pillars (W4-S9 lives in P32). Navigable by a machine, not by a reader. |
| **26,401–40,000 P32–P37+allocations** | **UNFIT AS IT STANDS** | Not for the content — X4's bridge, X7's recall and X9's Dala'il all re-derived exactly, and family E is well served here. **For the record.** 10,000 contiguous nodes across eight unrelated pillars carry one identical evidence string (F3), and inside that block sit **500 RESERVE-EARNED nodes sealed by a panel sitting where their own charter permits only a landing (F1).** A gift layer built over this band would publish 10,000 nodes that cannot be told apart and 500 that claim an acquisition that did not happen. **This band must be re-marked per-domain, and the 500 returned to reserved-awaiting-acquisition, before anything public sits on it.** |
| **40,001–61,600 P42 gap-fill** | **FIT WITH NOTES — the notes are load-bearing** | Its evidence *discipline* is the best in the waqf: every node names a `CARRIED BY` path, zero bare-prose nodes. Its content re-derived exactly (rows 10–12). But it is **0.39 KB per node on 104 files**, 900 of its nodes cite a script instead of an output, its creed layer drops the Tahawiyya's attribute articles, its ahkam floor misses ḥajj and zakāh, and its own doctrinal gate is blind to the biggest half of the corpus. **Publishable as a candidate-and-census layer, labelled as such. Not publishable as a finished body of knowledge.** |

**On the gift layer itself.** The proposal adds 41,600 nodes on top of 61,600. This bench measured
the direction of travel: **13,800 core nodes → 1,621 files. 21,600 P42 nodes → 104 files.** Node
count has been growing 9.4× faster than evidence for four bands running. A gift is judged by what a
stranger finds when he pulls on one thread. **Fix the record before adding to it** — the four
mechanical repairs (F1 re-mark 500, F2 lift the 3 MB exemption and the `[:3]` cap, F3 re-mark the
10,000-node block per-domain, F10 the `+3` offset) are all small, and together they change what the
whole 61,600 means.

**Never adjudicated.** No school was weighed anywhere above. Where the Tahawiyya's articles are
named it is to say they are *missing from a file*, never to rank a creed. Where fiqh verses are
counted it is to say a *term floor* under-recalled, never that a ruling is right. Family C's and
D's contested ground is documented here exactly as the charter requires: both sides named, neither
crowned, and the bench's only verdict is about **files**.

---

*Bench 2 · 167 machine-scholar seats · families C, D, E · MACHINE-SCHOLAR, labeled · documents,
never adjudicates · never a mufti · human review invited and never waited for.*

FULL-AUDIT BENCH 2 VERDICT: FIT WITH NOTES - every one of the twelve numbers this bench re-derived across all five bands came back exact, but the gift layer must wait on four mechanical repairs to the record, of which two are blocking: the 500 RESERVE-EARNED nodes sealed by a panel sitting where their own charter permits only a landing, and the 10,000 contiguous nodes across eight unrelated pillars that share one identical evidence string.