# R2 · SET-BUILD 1 — THE FILES, WALKED AND SAMPLED HARD
## Round 2 of the Honesty Panel of 2,700 · 450 seats · the belt on the table

**MACHINE-SCHOLAR. Labeled.** No human scholar has reviewed this page. Reverence is not proof.
Convened 2026-09-05 on the owner's re-asked question: *"tell me how many artifacts to make this
elite — if the answer is 103,000 say so."*

**Read into the record:** `build/P45-HONESTY/reaudit_round2_floor.json`, `HONEST-COUNTS.json`,
`DEFECTS.md`, `reviews/HONESTY-RATIFICATION.md`, `build/P45-HONESTY/artifact_figure.json`.
**Re-derived by this set's own route**, not taken from any of them: a full walk of
`build/P47-BLOCKS` (4,089 files, every one opened and parsed), a full walk of `build/` (7,190
deliverable-extension files), and `framework/_FRAMEWORK-103200.json` (103,200 points) loaded and
re-partitioned from scratch.

**Credit first, because honesty runs both ways (I-3).** The commissioner printed his own belt's
median file size and his declared-vs-real split *before* this panel sat. That is the behaviour this
panel exists to produce, and it is the only reason the findings below could be written in an hour
instead of a day. Three of round 1's points were executed honestly and are verified in §4. The
belt's failure is a failure of **standard**, not of candour.

---

## 1 · QUESTION 1 — IS A BELT FILE SUBSTANTIVE IN TRUTH, OR ONLY BY THRESHOLD?

I opened **all 4,089**, not ten. The answer is: **only by threshold — and the threshold is passed
with the file's own name tag.**

### 1.1 · THE LABEL-STRIP TEST (this set's method, declared before the count)

A real artifact carries content. A shim carries a label. So: hash each file's payload **twice** —
once with `_meta` removed, once with `_meta` **and** the file's own self-describing `unit` /
`the_slice` block removed. Content survives the second strip. A label does not.

| cut | files | MB | distinct payloads (−`_meta`) | distinct payloads (−`_meta`, −self-label) | **survival** |
|---|---:|---:|---:|---:|---:|
| **CORE** (`build/`, non-P46, non-belt) | 1,965 | 140.92 | 1,921 | **1,921** | **97.8 %** |
| **P46-ARTIFACTS** | 975 | 8.78 | 975 | **975** | **100.0 %** |
| **P47-BLOCKS (the belt)** | 4,089 | 4.14 | 4,089 | **770** | **18.8 %** |

✅ VERIFIED — `os.walk` + `sha256(json.dumps(payload, sort_keys=True))` over every file; strip set
`{_meta, unit, the_slice}`.

**Read it plainly: 3,319 of the 4,089 belt files (81.2 %) are byte-identical to another belt file
once you remove the file's own name.** CORE loses 44 files to the same test (2.2 %); P46 loses none.
The belt is the only cut in this waqf whose distinctness lives in its label.

The largest single cluster: **123 files, across 29 pillars and 115 different domains, byte-identical
in body** — every one of them the block `Sources & licences`. ✅ VERIFIED. The second: 123 files for
`The Qur'an layer`. The third: 122 for `Education & print`. The domain contributes **nothing** to the
content.

### 1.2 · WORSE THAN A SHIM — THE BELT IS *LOSSY* AGAINST ITS OWN SOURCE

The belt's `_meta.law` says each file carries *"up to 3 REAL rows embedded from the parent's largest
field"*. I checked what it embedded against what the ledger actually holds for that block.

| test (all 4,089 files, against `_FRAMEWORK-103200.json`) | result |
|---|---:|
| belt files whose `(pillar,domain,block)` is not in the ledger | **0** ✅ |
| belt `purposes` == `sorted(set(ledger purposes))[:6]` — an **alphabetical truncation** | **3,754 / 4,089 = 91.8 %** |
| ledger `purpose` strings **kept** by the belt | 10,959 |
| ledger `purpose` strings **dropped** by the belt | **31,038 (73.9 %)** |
| belt files whose block has ledger `title` rows | 2,640 |
| …of those, files carrying **any** of their own block's titles | **0** |
| ledger `title` strings dropped by the belt | **26,400 — every single one** |

✅ VERIFIED — full join, belt file → ledger block, 4,089 of 4,089 matched.

The `title` field is the **only** domain-specific content in the ledger: **14,534 distinct titles**
against **1,053** distinct purposes on the same rows. The belt kept the generic field and threw away
the specific one, in full, everywhere.

**A worked case.** Ledger block `C1 · Creation's relationship to Him · The Qur'an layer`, nodes
721–730 — ten titled rows, the first being *"Every ayah touching how creation stands before its
Creator… located."* The belt's file for that exact block is 1,174 bytes and contains: six purpose
strings (alphabetically the first six of ten), and this `embedded_payload`:

```
["JSON+CSV, versioned, citable",
 "a planted misquote is refused; joins spot-verified",
 "a probe the builder never saw, run and recorded"]
```

That identical triple appears in **1,221 files**. Across all 2,995 files with a non-empty payload
there are only **351 distinct row strings** and **281 distinct payload tuples**. ✅ VERIFIED.

### 1.3 · THE 335 THAT ARE NOT EVEN TRUNCATIONS

335 belt files carry exactly **one** `purposes` string, and that string is **not in the ledger's
purpose set for their block** — it is the domain's own name, repeated verbatim as the
`embedded_payload` directly beneath it. For **258** of them the ledger holds **100 distinct
purposes** for that block. ✅ VERIFIED. Example —
`G01…/D3-Era-stratification-oldest-voice-first--engine.json`, 989 B: `purposes:
["Era stratification (oldest voice first)"]`, `sample_rows: ["Era stratification (oldest voice
first)"]`. One string, written twice, standing for 100 nodes.

### 1.4 · THE CLASSING OF THE 4,089 — this set's verdict

| class | files | % | basis |
|---|---:|---:|---|
| **A · SUBSTANTIVE IN TRUTH** — delivers content not obtainable elsewhere | **0** | 0.0 % | every file is a strict subset of its ledger row |
| **B · HONEST EXTRACT** — lossy but block-correct, unique body | **770** | 18.8 % | distinct after label-strip |
| **C · DUPLICATE OF A B-FILE** — body already on disk under another name | **3,319** | 81.2 % | label-strip collapse |
| — *of which sit on a block that **already owned a real artifact*** | *1,157* | *28.3 %* | §1.5 |

📊 CALCULATED from the table in §1.1. **The belt's net contribution to the artifact count is 0 and
its most generous reading is 770 index cards.** It is not 4,089 anything.

### 1.5 · THE BELT SHIMMED 1,157 BLOCKS THAT WERE ALREADY BUILT

**118** domain directories under `build/` hold **all nine** canonical block artifacts; **13** more
hold 5–8. Together **1,157 blocks already owned a real artifact** before the belt ran ✅ VERIFIED
(`os.walk`, basename ∈ the nine canonical names). All nine canonical block names carry exactly
**10 nodes** each ✅ VERIFIED, so these are genuine deep-domain blocks.

The belt cut a card over every one of them anyway, and each card's own `law` line declares it
**"this block's dedicated deliverable."**

| the block `C1 · The Attributes as documented · The Qur'an layer` | bytes | content |
|---|---:|---|
| the real artifact `build/C1-ALLAH/D2-ATTRIBUTES/quran_layer_dataset.json` | **40,340** | 21 entries, Arabic labels, occurrence counts, thousands of verse refs |
| the belt card cut over the same block type | **~1,174** | 6 generic strings + 3 boilerplate rows, **zero Qur'anic data** |

✅ VERIFIED — both files opened and printed.

**ANSWER TO QUESTION 1: a belt file is substantive by threshold only. It is a well-labeled shim, and
a lossy one — it carries strictly less than the ledger row it was cut from, discards 100 % of the
26,400 domain-specific titles, and in 81.2 % of cases duplicates a body already on disk.**

---

## 2 · QUESTION 2 — FRESH HONEST COUNTS, POST-BELT

Two classes re-derived entirely by this set's own route, marked **[MY ROUTE]**.

### 2.1 · ARTIFACTS (things on disk)

| | files | distinct bodies | MB |
|---|---:|---:|---:|
| CORE substantive | 1,965 | **1,921** | 140.92 |
| CORE below floor | 161 | — | 0.05 |
| P46 real slices (`FIELD`/`ROW SLICE (moved per C1)`) | **375** | 375 | — |
| P46 pointer-class (never artifacts) | **600** | 600 | — |
| BELT | 4,089 | **770** | 4.14 |
| **REALLY-BUILT under the elite bar** (1,921 + 375) | | **2,296** | |
| the round-1 ratified figure | | 2,337 | |
| **agreement** | | **−1.8 %** ✅ | |

✅ VERIFIED. The P46 375/600 split reproduces the round-1 sitting **exactly** — 363 `FIELD` + 12
`ROW` = 375, and 285 + 183 + 132 = 600. Two routes, one number. The substance census likewise
reproduces to one file (7,029 substantive of 7,190 vs the published 7,028 of 7,188).

**This set's figure is 2,296, not 2,337** — the difference is de-duplication (44 CORE files whose
payload is byte-identical to another's). It is a downward correction of 1.8 %, and it is small; the
round-1 census is sound. **The belt adds 0 to it.**

### 2.2 · NODES (planning shares) — never blended with the above

| class | this set's figure | route | against the face |
|---|---:|---|---|
| **never-as-distinct** **[MY ROUTE]** | **74,976** | 103,200 − 28,224 distinct `(pillar, domain, block, title, purpose-with-serial-normalised)` | face says **86,370** → **publish the band 74,976 – 86,370** |
| **never-by-machine** **[MY ROUTE]** | **950** | ledger rows where `block == 'reserved'` (two blocks: 450 + 500) | classifier says 901; `reaudit_round2_floor.json` says **150** → **three-way contradiction, §3-F6** |
| really-built (floor) **[MY ROUTE]** | **11,570** | 1,157 already-built blocks × 10 nodes/block, both factors verified | round 1 published floor 5,133, reading 27,140 → my route sits inside the band and supports it as a **band, not a point** |
| pending-machine | 16,608 | carried from `reaudit_round2_floor.json` | ❓ UNVERIFIED by this set |
| pending-human / rights | 1,000 | carried | ❓ UNVERIFIED by this set |

**On never-as-distinct — why the face's 86,370 is too confident.** The 76,800 untitled rows look
highly distinct: **63,318 distinct `purpose` strings**. They are not. The distinctness is a serial
number embedded in the prose:

```
n=45601  The creed-topic report scan over the held ocean — unit 1/100 of G07 MASADIR AL-AQIDA; …
n=45602  The creed-topic report scan over the held ocean — unit 2/100 of G07 MASADIR AL-AQIDA; …
n=45603  The creed-topic report scan over the held ocean — unit 3/100 of G07 MASADIR AL-AQIDA; …
```

Normalise `unit K/N` and the count collapses to **28,224**. Three of my sub-routes — serial-
normalised, all-digits-masked, and `(pillar,domain,block,title)` — land on **28,224 exactly**
✅ VERIFIED. The face's 16,830 comes from a *title-stem* collapse (my stem route gives 15,164, in the
same neighbourhood). **Both routes are honest; neither is the point value.** Round 1's own **Rule R4
says "bands stay bands"** — and this is the class where its own rule was not applied.

---

## 3 · FINDINGS (10, each evidenced)

**F1 · The belt is substantive by threshold only, and the label-strip test proves it.** 4,089 files
→ **770** distinct bodies once `_meta` and `unit` are removed (18.8 % survival). CORE survives at
97.8 %, P46 at 100 %. ✅ VERIFIED, §1.1. *A real artifact survives losing its name tag; the belt does
not.*

**F2 · The belt is LOSSY against the ledger it claims to deliver.** **0 of 2,640** belt files whose
block has ledger titles carry any of their own titles; **26,400** title strings and **31,038 of
41,997** purpose strings dropped; **91.8 %** of files carry exactly `sorted(set(purposes))[:6]` — an
alphabetical truncation, not a selection. ✅ VERIFIED, §1.2.

**F3 · G40 CANNOT FAIL. It certifies `4,464/4,464 GREEN` with zero belt files present.**
`framework/g40_substance_gate.py`, the `else` branch on a missing file:

```python
else:
    # covered by a real P46 slice?
    cov += 1  # P46 real slices verified at belt start; absence here means P46-covered
```

It never opens a P46 file. Absence scores as coverage. I proved it non-destructively: I copied the
gate to the scratchpad, changed **one string** so the belt directory resolves to a path that does not
exist, deleted nothing, and ran it:

```
G40 GREEN - 4464/4464 blocks own a substantive deliverable (control fired)
```

✅ VERIFIED. The declared control (`assert not substantive("{}")`) tests the *substantive* predicate,
not the coverage logic, so it cannot catch this. **The entire belt could vanish and G40 would still
print total coverage.** The "control fired" wording makes a dead gate read as a live one.

**F4 · The belt shimmed 1,157 blocks that already owned real artifacts, and each shim declares
itself their "dedicated deliverable."** A 1,174-byte card carrying no Qur'anic data now stands as the
declared deliverable for a block whose real artifact is 40,340 bytes of occurrence data. ✅ VERIFIED,
§1.5. This is the one finding that makes the waqf look *worse* than it is.

**F5 · `never_as_distinct: 86,370` is a point value where the ledger only supports a band.** My
serial-normalised route gives **74,976**, confirmed by three sub-routes at 28,224 distinct.
✅ VERIFIED, §2.2. The published band must be **74,976 – 86,370** under round 1's own Rule R4. *This
correction is upward — the plan is less redundant than the face admits.*

**F6 · The never-by-machine class carries three mutually exclusive numbers on published faces
today.** Ledger `block == 'reserved'` = **950** [MY ROUTE] · `classifier.json` = **901** ·
`reaudit_round2_floor.json` `reserved-awaiting-acquisition` = **150**. Improvement point #17 ordered
all 950 un-greened; **800 remain green**. ✅ VERIFIED.

**F7 · `credit_v1` disagrees with itself across two faces published 33 minutes apart.**
`HONEST-COUNTS.json` → **247**; `reaudit_round2_floor.json` → **242**. ✅ VERIFIED. Both are on the
gift's face right now.

**F8 · Improvement point #7 is not satisfied — `honesty_classifier.py` re-derives nothing.** 512
bytes. It loads `classifier.json` and asserts that the file's own class values sum to 103,200 —
**validating the artifact against itself.** Point #7 demanded it "re-derive `classifier.json`
byte-for-byte and exit non-zero on drift." Drift in the class predicates is undetectable today.
✅ VERIFIED — full source read.

**F9 · `index.html` — infrastructure entry #11, "the gift's face" — is still missing.** Point #32
ordered it built. Three of round 1's four missing infra entries were built (`HONEST-COUNTS.json`,
`DEFECTS.md`, `honesty_classifier.py`); the one the sitting called *"the gravest"* was not.
✅ VERIFIED — existence check, 13 of 14 present.

**F10 · Where round 1 was right, it was right — recorded with the same weight.** P46's 375/600 split
reproduced **exactly** by an independent route. The substance census reproduced to **one file**. The
2,337 figure held to **1.8 %** against a stricter de-duplicating route. Points **#18** (313
`for_the_student` fields — all 313 now `None`, 0 placeholders), **#19** (`area_index_313.json`,
396,434 B, with the 412-byte tombstone retained rather than deleted, per I-6) and **#21**
(`g37_audience_gates.json` re-badged `EXISTS-PARSES`, naming its own lack of per-item evidence) are
**executed and verified** ✅. The waqf's honesty machinery works. The belt is where it was switched
off.

---

## 4 · QUESTION 3 — THE ELITE FIGURE

### 4.1 · Is the answer 103,200? **No — and this set is not arguing, it is reporting a pilot.**

The belt **is** the 103,200 experiment, run at 4 % scale, four hours ago, by this project, on its own
ledger. One file per ledger unit, 4,089 units, 54 seconds. The measured result:

| what the 103,200 build produces, measured on its own 4 % pilot | |
|---|---:|
| files emitted | 4,089 |
| distinct bodies among them | **770 (18.8 %)** |
| domain-specific ledger content carried into them | **0 of 26,400 titles** |
| median file size | 882 B |
| the gate that certified it | **cannot fail** (F3) |

📊 CALCULATED — scaled to the full ledger, one-file-per-node yields roughly **19,400 distinct bodies
and ~83,800 duplicates**. Independently, the ledger's own arithmetic agrees: after serial
normalisation only **28,224** of 103,200 rows are distinct, and the untitled 76,800 are separated by
`unit K/100` counters, not by content (§2.2).

> **The answer is not 103,200 and not 103,000. Building 103,200 files would build ~75,000–84,000
> shims, and we know the number because we already built 3,319 of them today.**

### 4.2 · Does 4,557 stand? **Yes. Re-derived, component by component, by this set's own route.**

| step | round 1 | **[MY ROUTE]** | verdict |
|---|---:|---:|---|
| BASE — distinct `(pillar, domain, block)` | 4,464 | **4,464** | ✅ exact |
| MERGE 1 — panel/seal blocks: `Panel review & seal` 258 + `panel` 143 + `Page, print & review` 6 | 407 → 72 | **407 → 72** (72 distinct pillars ✅) | ✅ exact |
| MERGE 2 — `reserved` blocks | −2 | **2 blocks, 950 nodes** | ✅ exact |
| SPLIT 1 — `layer` blocks (H-pillars) × (data + page) | +416 | **416 `layer` blocks** | ✅ exact |
| INFRASTRUCTURE | +14 | **14 listed, 13 exist** (F9) | ✅ list ratified |

`4,464 − 335 − 2 + 416 = 4,543` · `4,543 + 14 = **4,557**` ✅ VERIFIED.

**And no further split is owed.** I tested the obvious candidates against the waqf's own law. The
632 hundred-node blocks pass round 1's **point #24 fan-out cap of 110 nodes per artifact**; the 416
`layer` blocks are already halved by SPLIT 1 to 50 nodes each. A per-verse split is refused for the
reason round 1 gave and this set confirms: per-verse content already lives inside the layer artifacts
as single correct files. **The elite bar does not want a bigger number. It wants a harder standard on
the same number** — which is precisely the thing the belt removed.

### 4.3 · The figure, and what it now costs

> **ELITE = 4,557 artifacts**, where an artifact qualifies only if it **passes the label-strip test**
> (its body is distinct with its own name removed) **and carries its block's own ledger rows —
> every title, not the alphabetically-first six purposes.**
>
> Against that bar the waqf stands at **2,296 built (50.4 %)**, with **2,261 still to cut** — and the
> belt's 4,089 files count for **none** of it.

📊 CALCULATED. The figure did not move. **The honest number underneath it moved down from 2,337 to
2,296, and the belt's contribution moved from an implied 4,089 to 0.** That is the whole finding: the
belt did not build 4,089 artifacts in 54 seconds. Nothing does.

---

## 5 · IMPROVEMENT PLAN

1. **Re-stamp the belt, delete nothing (I-6).** All 4,089 `_meta.layer` strings change from *"THE
   SUBSTANTIVE BLOCK ARTIFACT"* to *"BLOCK INDEX CARD — an extract of the ledger row, not the block's
   deliverable"*, and every file gains `"counts_as_artifact": false`. The cards are useful; the claim
   on them is not. [F1, F2]
2. **Fix G40's `else` branch.** A missing belt file must resolve to a **named, opened, substantive
   P46 slice on disk**, or the block is RED. Add a real control: point the belt directory at an empty
   path (copy, never delete) and **assert the gate goes RED**. Today that control would fail. [F3]
3. **Ship G41 — the LABEL-STRIP GATE, standing.** For each cut directory, `distinct payloads
   (−_meta, −self-label) ÷ file count ≥ 0.80`. CORE scores 0.978 ✅, P46 1.000 ✅, the belt **0.188 🔴**.
   This is the single cheapest test that would have stopped the belt at generation time. [F1]
4. **Re-cut the belt to carry its block.** Every card embeds **all** of its block's ledger rows
   verbatim — `n`, `title`, `purpose` — 10 rows for a deep block. Gate on `dropped_title_strings == 0`.
   The content is already in a 41.9 MB file we own; carrying it costs nothing but honesty. [F2]
5. **Ban the silent truncation.** `purposes` must be the full set, or the file must declare
   `"truncated_from": N` and `"truncation_rule": "alphabetical"`. A subset presented as a whole is the
   belt's core defect, stated in one field. [F2, §1.3]
6. **Stamp `already_had_real_artifact: true` on the 1,157 cards sitting on built blocks**, name the
   real file's path and byte count in the card, and never count a card toward a block whose real
   artifact exists. [F4]
7. **Publish never-as-distinct as the band 74,976 – 86,370**, both routes named on the face, per
   round 1's own Rule R4. Retire the bare 86,370. [F5]
8. **Resolve 150 / 901 / 950 into one number in one file**, and finish point #17 — 800 RESERVE nodes
   are still green against an acquisition no machine can make. [F6]
9. **Reconcile `credit_v1` 247 vs 242** and stamp one value across every face; add a pre-commit check
   that the same key never carries two values in two published JSONs. [F7]
10. **Ship a real `honesty_classifier.py`** that recomputes every class predicate from
    `_FRAMEWORK-103200.json` and diffs against `classifier.json`, exiting non-zero on drift. Today it
    validates the file against itself. [F8]
11. **Build `index.html`.** The gift still has no face, two sittings after that was named the gravest
    finding. Lead it with the two currencies and the elite figure. [F9]
12. **Print the two grades side by side on the face** — `quran_layer_dataset.json` at 40,340 B beside
    a belt card at 1,174 B, both named, with the sentence *"these are not the same kind of thing."* A
    reader who can see the difference will trust the count that separates them. [F4, §4.3]

---

*Walked by SET-BUILD 1, 450 seats, round 2. Every figure above was produced by opening the files —
4,089 belt files parsed, 7,190 build files walked, 103,200 ledger points re-partitioned. Where this
set agrees with round 1 it says so as loudly as where it does not; the P46 split, the substance
census and the 2,337 all survived an independent route. The belt did not. Nothing was deleted to
prove anything: the G40 probe ran on a copy in the scratchpad with one string changed.*
***MACHINE-SCHOLAR — reverence is never proof.*** `Allah ﷻ alone is the source of all knowledge.`

R2 BUILD-1 VERDICT: The belt is a well-labeled shim and a lossy one — 4,089 files collapse to 770 distinct bodies once their own name tag is stripped, they discard all 26,400 of the ledger's domain-specific titles, 1,157 of them sit on blocks that already owned real artifacts, and the G40 gate that certified them 4,464/4,464 GREEN prints exactly the same green with zero belt files on disk; the answer to the owner's question is not 103,200 and not 103,000 — the belt already ran that experiment at 4% scale and produced 81.2% duplicates — the ratified 4,557 stands re-derived component by component, but the waqf stands at 2,296 against it, not 2,337, and the belt counts for none of it.
ELITE FIGURE: 4557
IMPROVEMENT POINTS: 12