# RUN-001 — corpus identity: the forty-two curated documents are the archived originals

**Arm:** 懸衡 (Xuán Héng) / AMRAL Research Lab
**Date:** 2026-09-24
**Subject:** the Lebesgue case as published at [amral.evemisslab.com/lebesgue/](https://amral.evemisslab.com/lebesgue/) (site revision `ba14e814`; documents last changed at `9f3a3d76`, 2026-09-22), against the two source zips archived on `agent/lebesgue-finite-closure`
**Tools:** [`corpus.py`](../code/corpus.py), [`fetch_source.py`](../code/fetch_source.py), [`run001_corpus_identity.py`](../code/run001_corpus_identity.py)
**Logs:** [`run001-corpus-identity.json`](../data/gate-logs/run001-corpus-identity.json), [`SOURCE_MANIFEST.v1.json`](../data/SOURCE_MANIFEST.v1.json)

**Verdict: PASS.** All 42 curated documents are byte-identical to an archived
original, each under its own original filename, each bound by exactly one page,
and no two pages share a document. The hub page's sentence about its own files
holds as written. The drill planted 220 defects; every one was caught by exactly
the checks named for it and by no other.

---

## Why this is round one

The hub page says of its file list:

> 以下 42 篇是本頁全部頁面各自對應的原始文件,未經改寫。

*(The following 42 are the original documents behind each of this page's pages,
unrewritten.)*

Every later round reads a curated document and reasons about it as though it
were the archived original. If that were false anywhere, every later round
would inherit the gap without noticing. So it is settled first, and settled the
only way that can settle it: **by content hash, never by filename.** A filename
match cannot see a curation pipeline that normalised line endings, added a byte
order mark, or trimmed a final newline — and on the BSD line the same check did
find a curated document whose bytes had changed on the way in
(`bsd-verification-zhuiheng`, RUN-001). Here none has.

## What was compared

| side | what | count |
| --- | --- | ---: |
| curated | the site's `lebesgue/files/*.md` | **42** |
| archived | every leaf of both zips, nested zips walked in memory | **1,459** (126 of them `.md`) |

| | |
| --- | ---: |
| curated with a byte-identical archived original | **42** |
| of those, renamed on the way in | 0 |
| of those, with more than one identical archived copy | 0 |
| curated with no byte-identical original | **0** |
| archived `.md` not curated | 84 |

The 84 are not a finding. The site presents a reading path; the packages also
carry READMEs, state crystals and verification-pass notes. Only the other
direction is a claim about fidelity.

Where the 42 originals sit:

| group | curated | original in |
| --- | ---: | --- |
| Round 00, methodology freeze v0.2 | 1 | main archive, top level |
| Rounds 01–36 | 36 | main archive, one per round package |
| Round 37 | 1 | the corrected standalone Round 37 package |
| DLMVC · BCODR · LESR (Paper 00) · UESFCM | 4 | main archive, one per methodology folder |

Round 37 is the one place provenance could have gone wrong: the main archive's
own Round 37 entry is a bare cell-49 certificate with no narrative document, and
the narrative arrived separately in a corrected package. The curated Round 37
document is byte-identical to the corrected package's copy.

## The checks

| id | guards | result |
| --- | --- | --- |
| C1 | both source zips match their pinned size, SHA-256 and git blob id | PASS |
| C2 | every curated document has a byte-identical archived leaf | PASS — 42/42 |
| C3 | `files/` holds exactly the documents the hub links, and as many as the hub's own text states (「以下 42 篇」) | PASS |
| C4 | the archive's full leaf manifest equals the committed one | PASS — written this round, 1,459 leaves |
| C5 | every page under `p/` references exactly one `files/*.md`, it exists, no two pages share one | PASS — 42 pages ↔ 42 documents |

C4 has no teeth in the round that writes the manifest; it exists for every
later round. From here on, any change to any of the 1,459 archived leaves — a
re-upload, a re-zip, a replaced package — trips it before anything else is read.

## The drill

A plant passes only when **exactly** its named set of checks trips: a named
check that stays green is blind, and any other check that trips fired at the
wrong scope.

| kind | cases | named checks |
| --- | ---: | --- |
| one byte flipped, per curated document | 42 | C2 |
| line endings converted, per curated document | 42 | C2 |
| UTF-8 byte order mark prepended, per curated document | 42 | C2 |
| final newline added or removed, per curated document | 42 | C2 |
| the archived original removed, per curated document | 42 | C2 + C4 |
| a curated document renamed, bytes kept | 1 | C3 + C5 (C2 must stay green: it matches content, not names) |
| an uncurated archived leaf changed | 1 | C4 only |
| a curated document deleted | 1 | C3 + C5 |
| an unlisted extra document added | 1 | C3 + C5 |
| a page re-pointed at another page's document | 1 | C5 |
| a page given two documents | 1 | C5 |
| a byte appended to a source zip | 1 | C1 |
| a pin edited | 1 | C1 + C4 |
| a committed manifest entry edited | 1 | C4 |
| the hub's stated count disagreeing with the files | 1 | C3 |

**220 / 220 plants caught by exactly the named checks.** Controls — nothing may
trip: the unmodified corpus; the archive's leaf order reversed; the insertion
order of documents, pages and hub links reversed. **3 / 3 undisturbed.** Across
all 223 cases, **849 / 849** assertions that a check outside the named set stays
green held.

The drill was itself drilled: each check in turn was replaced by one that
always passes, and the drill had to report it blind.

| check blinded | cases the drill reported blind | plants naming that check |
| --- | ---: | ---: |
| C1 | 2 | 2 |
| C2 | 210 | 210 |
| C3 | 4 | 4 |
| C4 | 45 | 45 |
| C5 | 5 | 5 |

Every check is load-bearing, and no plant is caught only by accident.

## Two notes that are not findings

**Round 00 exists in two versions; the site serves v0.2.** Between v0.1 and v0.2
the only changes are the version line and §26 (per-round output, an append-only
state crystal per round, and a two-conversation working arrangement). §1–25 and
§27–35 are identical. A later round that reads v0.2 reads the same mathematics
v0.1 froze.

**The hub's statement about external literature is not re-checked here.** The
hub says the four papers it cites were checked against their arXiv originals,
and places the published interval at `0.8344 ≤ a_Leb ≤ 0.8440935944`. That is
curation, not this round's subject. It will be checked in the round that first
depends on it — in particular the lower bound, since the line's own target is
measured against it.

## What this round does not establish

Anything about the mathematics. RUN-001 shows that what the site shows is what
the archive holds; it says nothing about whether what the archive holds is
right. That starts with RUN-002, Round 00.