Status: CAMPAIGN ENDED. Cleansing 107 inherited works was abandoned for acquiring fifteen correct ones — see ARCHIVE-CANON-SPEC, which supersedes this as policy. What survives here is the defect vocabulary, which is still the regression suite: a detector named below must keep reporting zero. Scale: 107 shelved works Prior art: scripts/audit-text-quality.mjs, which already exists and already knows this defect class.

R.I.S.E. shelves texts it did not typeset. Scans carry an editorial apparatus, and OCR folds that apparatus into the prose. Three findings, all real:

• Hamlet was not Hamlet. It was the Cambridge variorum, and the sigla came through as text: 140. at] Ff. om. Qq. Withdrawn.

• Three Shakespeares withheld behind the same fault, still awaiting an edition decision.

• Vitruvius, Book I leaks marginalia one atom at a time: ATHENS], ROME], Giocondo, Venice, 1511)]. These are not merely ugly — the Page's heading heuristic promoted two of them into section headings before it was tightened, which is what damage does when nothing filters it.

• Running heads mid-sentence — DECIDED: trim. See §2b, which is the most actionable finding in this document and the one with the clearest evidence.

• Stranded running heads, found 2026-08-04 while fixing a division boundary. Romance of the Three Kingdoms ends chapter CV with DEFEAT AND DEATH OF KUNGSUN YUAN; PRETENDED ILLNESS OF SSUMA I — the title of chapter CVI, which then opens correctly with its own CHAPTER CVI. / Defeat and Death of Kungsun Yüan; Pretended Illness of Ssŭma I. So a reading finishes by announcing the next one, and the announcement is duplicated verbatim at the head of that next one.

• The Mahabharata is partly the First World War. Found 2026-08-04 by the reviewer's first adversarial batch. Its last five "volumes" are the New York Times Current History of the European War (1915) — Kipling, Alfred Noyes, Sir John French's dispatches, the Battle of the Suez Canal. See §1b. This is the Hamlet fault again, and it is on the shelf right now.

• Anna Karenina, Chapter 1 begins with the book's navigation apparatus. Found 2026-08-11 while testing adjacent Workshop visual spans. The served chapter begins with title/author/translator matter, a literal Contents, and links for PART ONE through PART EIGHT before the genuine PART ONE / Chapter 1 opening. See §2g. This is a cleansing defect even though the runtime boundary failure it exposed was corrected independently.

The rule this scope serves: no such text in any book.

Status: TRIMMED 2026-08-05 (e976c03) — option 1, the interim. The 403,208 words are withdrawn: volumes 5–9 are gone, and the payload now holds volumes 1–4 with zero war vocabulary and its own intact (2,948 Arjuna, 2,913 Pandava, 2,067 Krishna). A reader can no longer be handed a 1915 war periodical labelled as Vyasa.

The acquisition is still open. Option 2 — re-sourcing the true remaining volumes — remains the right end state and needs a human; until then the dossier must say the edition is volumes 1–4. Everything below is the evidence the trim was made on, kept so the decision is not rediscovered.

Zero. Across four hundred thousand words presented as the Mahābhārata, not one occurrence of Bharata, Kuru, Pandava, Arjuna, Krishna, Yudhishthira, Parva, Rishi or Brahmana. What is there instead:

CHAPTER I. / THE FRENCH SETBACKS IN AUGUST. PART I. / THE CONDUCT OF THE GERMAN TROOPS IN BELGIUM. MUSIC OF WAR — By Rudyard Kipling

That is 13.7% of the work, and because those volumes were never divided, "Volume 5 — Front matter" is a single reading of 115,312 words. A reader entering there is handed a 1915 war periodical labelled as Vyasa.

The provenance record declares nine artifacts, #15474–#15482, each labelled "Project Gutenberg Ganguli volume N". The ingest assumed a contiguous identifier range. Ganguli's Mahābhārata occupies #15474–#15477; what follows in Gutenberg's numbering is a different work entirely. The labels are the ingest's own assertion, not something read back from the files — so the dossier says "Ganguli volume 5" about a text that says "Kaiser".

identity.test.js asks whether a work names itself in its own pages, and the Mahābhārata does, seventeen thousand times, in volumes 1–4. The check is made of the WORK and the reading is served at the DIVISION. A work can pass its identity test while five of its divisions are a different book.

The guard this suggests — and it is cheap, because the signal is not subtle: no division may be devoid of the vocabulary its work is dense in. Volumes 5–9 score zero against a corpus averaging one proper noun per 145 words. A per-division identity check is rung 2 of §3 in its simplest possible form.

BUILT 2026-08-12. scripts/audit-division-identity.mjs derives it and src/content/archive/division-identity.test.js holds it. See §3b.

Eight shelved works are built from multiple artifacts; only two assume a sequential identifier range. The other is Le Morte d'Arthur (#1251–#1252), and both its volumes are genuinely Arthurian — 4,893 and 5,464 hits. This is one work, not a class.

1. Trim volumes 5–9 and keep the 2.55M words of genuine Ganguli. The deletion is well-evidenced and leaves the work honest but incomplete — and the dossier must then say the edition is volumes 1–4.
2. Re-source the true remaining volumes and keep the work whole. Correct, and it is an acquisition, not an edit.
3. Withhold entirely. Disproportionate: four volumes are sound.

The recommendation is 2, with 1 as the interim — a shelf should not serve the First World War as the Mahābhārata while an acquisition is pending.

Naming them is most of the work, because each has a different detector and a different disposition.

The last two rows already have machinery: identity.test.js proves a work names itself in its own pages, and divisions.js drops contents pages. Everything above them is open.

Decided 2026-08-04, on Mateo's call: remove them. This section exists because the class earns a decision the others do not yet have, and because it is the one place where deletion is provably safe rather than merely reasonable.

A reader met this on the Page and read it, correctly, as a visual glitch:

…the men were standing, rifle in hand, their eyes fixed on the ground in front of them. Now and GUILLEMONT 101 then by the light of a rocket I saw the gleam of helmet after helmet…

GUILLEMONT 101 is not a heading, a title, or a seam between movements. It is the running head of the printed page — the chapter name and the recto page number — and OCR left it wherever the physical page turned, which is normally mid-clause. The Storm of Steel carries the sequence 93, 95, 97, 99, 101 … 109.

The raw span in the payload shows both halves of the furniture, the verso numeral and the recto head:

…One could see that the man had been \n\n92 \n\n\nGUILLEMONT 93 \n\n\nthrough horror to the limit of despair…

Delete the whole span and the sentence is restored, not edited: "…the man had been through horror to the limit of despair…". That is the test this disposition must meet.

Swept over all shelved payloads, 2026-08-04:

The heaviest carriers are the Shahnama (1,172 — KAI KHUSRAU×220, GUSHTASP×124), the Corpus Hermeticum (500), and the Storm of Steel (139, of which 110 are mid-sentence — the worst ratio on the shelf, which is why it surfaced first).

Trim a running head only where its POSITION proves it is furniture.

This is the same evidence canon R11 uses to refuse promoting one to a heading, applied one layer down to the text itself. A line qualifies when all hold:

1. It is a short line of capitals ending in an arabic number, and its capitalised stem repeats three or more times across the work with different numbers.
2. The preceding non-empty line does not end a sentence.
3. The following non-empty line resumes in lower case.

Conditions 2 and 3 are what make this a deletion rather than a guess. The other 1,576 occurrences are suspicions, not verdicts, and go to a reviewer — because a repeated capitalised stem followed by a number is also exactly the shape of a legitimate division heading. The sweep found BOOK×18 in the Mahabharata and CH. III. INTRODUCTION.×5 in the I Ching, and a rule that trimmed those would delete the work's own structure. That is the entire reason this section is narrow.

Bare numerals are explicitly NOT in scope for automatic trimming. 4,531 of them exist and many are the work's own structure — a hundred verses from Old Japan numbers its poems, King Lear's scan numbers its lines. Only a bare numeral inside a qualifying furniture span (the 92 above, adjacent to the head it belongs with) is removed, and only as part of that span.

The span is the whole furniture run, not the head alone. A printed opening leaves both numbers behind — the verso page number and then the recto running head:

…set up an observation post. \n130\n\n\nIN THE VILLAGE OF FRESNOY 131\n\n\nI took a few men…

Spanning only IN THE VILLAGE OF FRESNOY 131 deletes half the furniture and leaves a naked 130 standing in the prose, which looks exactly as broken as leaving all of it. The span therefore runs from the end of the last real line of text to the start of the next one, swallowing the blank lines and any adjacent bare numeral.

Found by the reviewer. The first job builder spanned the head alone; the reviewer refused five of thirty with span: "too_small" and was right. This section had already said the numeral belongs in the span and the code had not. That is the span mechanism paying for itself on its first batch.

What replaces the span depends on whether a sentence was interrupted.

• The preceding text does not end a sentence → replace with a single space, so the clause rejoins.
• It does → replace with a paragraph break. We know the furniture was inserted; we do not know that the paragraphs on either side of it were ever one, and joining them would be an edit rather than a deletion.

This test asks less than the positional proof does, deliberately. Proof requires a lower-case continuation because it licenses deletion with no reviewer at all. Rejoining only asks whether the sentence had ended — "…furnished by the canteen at / 12 / FROM BAZANCOURT TO HATTONCHATEL 13 / Montcornet." is plainly one sentence, and the capital is a place name, not a new paragraph.

On Storm of Steel's thirty reviewable candidates: 7 rejoin, 23 keep the break.

The verification is a strict-deletion check, not a diff review: the result must be the original text with only the identified spans removed — the remaining characters, in order, must equal the original minus those spans. Anything else is a rewrite, and a rewrite here means the edition no longer describes what is on the shelf (§4's hard rule).

Storm of Steel first: worst ratio, smallest volume, and it is the text a Journey reads, so the fix is visible immediately in the Demonstration. Then the Shahnama and the Corpus Hermeticum, which are the volume.

Asked for by Mateo, 2026-08-05: all-capital words, punctuation in series like ;;;...., and brackets. Swept exhaustively over 1,710,281 lines of 91 works (a one-shot audit script, since deleted).

The headline is that the counts are misleading, and acting on them would have been the largest mistake of this pass. Three of the four biggest buckets are the edition doing its job.

205 bare [Illustration] markers. They stand for a plate the printed edition had and this one does not; rendered, the reader is shown those characters. That is a broken frame written in words, and reverent degradation is the standing rule: a work that will not resolve is absent, never a broken frame.

Only the bare ones. [Illustration: “I'm the tallest”] carries a caption. And [Illustration] BUTTERFLY DANCE has its caption outside the bracket — removing that marker would strand an all-capital line between two paragraphs, which the compositor then reads as a title. Removing furniture in a way that manufactures a heading is not an improvement; it is R11's fault arriving by another door.

CORRECTED 2026-08-05, and the correction matters more than the finding.

The line above used to read: "The orphaned-bracket detector is wrong, and 7,223 counts prove it… The real orphans are rare." That was wrong, and ATHENS] — the case named in §1 from the very beginning, and pointed at twice — was sitting inside that dismissed bucket the whole time.

Two mistakes produced it, and both are about the instrument rather than the corpus:

1. The samples were not samples. readdirSync returns alphabetical order and the bucket kept the first N examples it saw, so every illustration in that survey came from a-doll-s-house — a stage play, the worst possible representative for a question about brackets. I read four lines from one work and wrote off 7,223 findings across 39. A sample that cannot show you a second work cannot tell you what a bucket holds.
2. The caps detector excluded the shape the question was about. It returned early on any line with no lower-case letter, reasoning that the compositor already judges headings (canon R11). But R11 decides how to set a line, not whether it belongs in the book. ATHENS] is a standalone all-capital line.

Both are fixed: samples now spread one-per-work before any work gets a second, the bracket test carries open-bracket state across lines so a stage direction's closing line is no longer an orphan, and standalone capitals are counted rather than skipped.

The corrected measurement. Standalone orphaned-] lines, nothing open, ≤60 characters:

The class is not one class, which is why no corpus-wide rule fits it:

• Vitruvius, 33 — plate captions and credits: ATHENS], ROME], EXAMPLE OF OPUS INCERTUM. THE CIRCULAR TEMPLE AT TIVOLI], (From his edition of Vitruvius, Venice, 1511)]. Trim — see §2e.
• Crane 86, Dresser 1 — figure caption tails, same shape.
• The Little Clay Cart 80 — P. 4.7], footnote anchors, not captions.
• Pride and Prejudice 34 — Chapter I.] and a repeated copyright line.
• And the trap: metamorphoses has in a]l and And al] the ground — the ] is a misread letter. Those are excluded by the standalone rule, and they must be: removing that bracket would be a repair, which §4 forbids outright. Flag, never touch.

[Greek: taxis] is not removable by deletion. It is the transcriber's substitute for Greek characters the plain-text format could not carry, and the Vitruvius page reads "Order (in Greek [Greek: taxis])". Deleting the bracket leaves "in Greek )"; deleting the parenthetical removes the Greek term the translator put there. Turning it into taxis would be a rewrite, which §4 forbids outright. This one is a rendering problem, not a text problem — the Page could present it as a gloss — and it is recorded here rather than acted on.

Scanned 2026-08-05 with npm run audit:text, which already existed and is the instrument that found Hamlet. The question was whether a large-grain defect is still on the shelf; the answer is no.

0 works score above 12 (likely wrong edition or bad scan)
0 more between 6 and 12 (worth a human look)
88 appear clean

Hamlet scored 32.3% apparatus and is withheld. The worst work now on the shelf scores 2.8, and its components are fine-grained OCR noise rather than a wrong edition:

The one apparatus signal above 1% was checked and is not a defect. Doctor Faustus's 35 apparatus-shaped lines are Alexander Dyce's own footnotes — [Footnote 32: concise syllogisms-- Old ed. "Consissylogismes."] — which is the edition doing its job, not OCR folding a variorum into the play. It is the [Footnote: …] class of §2c: keep, flag.

Justified: stopping here. There is no Hamlet-scale problem hiding, so building detectors for apparatus, marginalia, imprint fragments and OCR corruption would be hunting a 1–3% residue that no reader has complained about, in preference to work that has been waiting.

Not justified: calling those classes closed. They are unscanned at the instance level — this is a per-work score over a 240,000-character sample, not an exhaustive sweep. It answers "is a whole book wrong", which is the question that matters, and not "how many stray sigla exist", which is the question that does not.

The residue is recorded rather than pursued.

The survey above was made with the broken sampler, so it was re-run once the sampling and bracket-state faults were fixed (§2c). With samples spanning many works instead of one play, the large buckets hold up: caps-standalone is ACT I., BEOWULF, ANNA KARENINA — the works' own titles; shout-short is IOU and stage directions; bracket-number is footnote markers and their notes. No second Vitruvius was hiding in them.

What remains genuinely open is bounded and named: 244 standalone orphaned-] lines across 12 shelved works whose lines have not been read — Crane 86, the Little Clay Cart 80, Pride and Prejudice 34, the Shahnama 21, and eight works with fewer than ten each. The machinery to close them exists and is verified; closing them is a reading task, not an engineering one. A work joins ORPHAN_CAPTION_WORKS after someone reads its lines, never before.

Cheapest first, and nothing reaches a reviewer that a regex could have settled — that is the whole economy of this.

Rung 1 — deterministic. Pure functions over the payload, unit-tested against known-bad fixtures drawn from the findings above. Unbalanced brackets, sigla patterns, bare numerals, repeated running heads, mojibake ranges. Fast, free, and it will catch the large majority because scanning artefacts are formulaic.

Rung 2 — statistical. Per-work outlier detection: a paragraph whose character profile departs sharply from the work's own norm (capital ratio, punctuation density, mean token length). This is what finds a defect nobody anticipated, and it produces suspicions, not verdicts.

Rung 3 — read it. Only what survives rungs 1 and 2, and only in the neighbourhood of the suspicion — never a whole book. This is the handoff.

Large volumes of prose reviewed by a cheaper model (Gemini 3.6 Flash or similar), one bounded task at a time.

Input — a JSON job per suspect passage:

{
  "workId": "vitruvius-ten-books",
  "edition": "Morris Hicky Morgan, Harvard University Press, 1914",
  "locator": { "division": "Book I", "charStart": 18432, "charEnd": 18510 },
  "before": "…two paragraphs of surrounding prose…",
  "passage": "ATHENS]",
  "after":  "…two paragraphs of surrounding prose…",
  "suspicion": "orphaned-bracket"
}

Output — a verdict, and nothing else:

{
  "workId": "vitruvius-ten-books",
  "locator": { "…": "…" },
  "verdict": "apparatus",
  "confidence": "high",
  "disposition": "trim",
  "note": "Marginal gloss naming the city discussed; not authorial prose."
}

A reviewer FLAGS. It never rewrites.

No model — and no script — emits replacement prose into the Archive. Every text here is public domain with its basis recorded per EDITION, and a silently "corrected" text is no longer the edition it claims to be. That is the same failure as Hamlet, arrived at from the opposite direction: a book presented under a name that no longer describes it.

Permitted dispositions are therefore a closed set:

• keep — a false positive; record it so the detector can be narrowed.
• trim — delete a bounded span that is demonstrably not the work. Deletion only. Recorded with its offsets in the work's dossier.
• withhold — the defect is structural; the work leaves the shelf and the shelf says why (the Shakespeare precedent).
• re-source — a better edition exists; withhold and queue the acquisition.

The prompt itself is written: CORPUS-REVIEWER-PROMPT.md, with the system prompt to send verbatim, the job envelope, six worked examples drawn from the actual shelf, and the batch discipline. Two things there refine this section: the reviewer's vocabulary deliberately excludes withhold and re-source (those need a human, per §7.3), and it reports whether the proposed span is exact — an inexact span forces keep, so a bad boundary is reported rather than acted on.

The judgement is genuinely easy — "is this line part of the book?" — and the volume is genuinely large. That is exactly the shape of task to delegate. The expensive part is the contract, and it is written down above precisely so the reviewer's job stays narrow enough that a small model does it reliably.

Sampling discipline: a stratified sample of every reviewer batch is re-read independently before its verdicts are applied. A reviewer that drifts is a reviewer that quietly deletes prose, and by construction nobody would notice.

Every applied verdict lands in the work's dossier alongside its rights basis: what was removed, from where, on whose judgement, and when. A trim with no record is indistinguishable from corruption a year from now.

Withheld works follow the existing precedent: absent from the shelf, with the reason stated where a reader would look — reverent degradation applied to the catalogue rather than to imagery.

0. Running heads (§2b) — decided, specified, and independent of everything below it. It has its own detector, its own verification, and 293 instances that need no reviewer. Worth doing first precisely because it does not wait on the ladder, the handoff, or any open decision.
1. Fixtures. Harvest known-bad passages from Hamlet, the three Shakespeares, Vitruvius and the Storm of Steel into a test corpus. Nothing is built before there is something to fail against.
2. Rung 1 detectors + unit tests. Report only; no mutation.
3. Run the report over all 107 and read the shape of it. The distribution decides how much of rung 2 is worth building.
4. Rung 2 outlier pass, if the report says it is needed.
5. The handoff harness — job generation, batching, verdict ingestion, sampling audit.
6. Apply, dossier-recorded, work by work, most-defective first.
7. A standing test: the fixtures never come back. This is the guard that keeps the shelf clean after the pass ends.

0. Stranded running heads — a trim with its own evidence. The case above is unusually safe to act on, and worth deciding separately from the rest. The proposed detector: a division's content ends with a heading run whose normalised text is repeated as the opening heading of the NEXT division. Only the duplicate is dropped; nothing unique to the text is touched, and the text itself supplies the proof by saying it twice. The precedent is already in divisions.js, which refuses to serve a contents page as a reading — this is the same act, smaller.

   It is still a deletion, so it is your call, not the splitter's. It was deliberately NOT bundled with the boundary fix: that fix only moves material, and this one removes it.

1. The Shakespeare edition — still blocking re-source. Moby remains the recommendation.

2. Trim threshold — does a single orphaned bracket justify touching a text at all, or is the bar "a reader would notice"? I would set it at a reader would notice, and let rung 1 report the rest without acting.

3. Who ratifies a withhold — a model's verdict alone should not empty a shelf. Suggest: withhold and re-source always require a human.

Found by a reader, from one character. A hermetic text served 2v📄 mid-page: a page image in the source scan had been transcribed as U+1F4C4 PAGE FACING UP. The question "how many more of these are there" turned out to have a bounded, useful answer.

The sweep: every codepoint above U+2100 in Unicode category So, Sk, Cn or Co, across all 91 work payloads. 22 distinct codepoints. They sort into three dispositions and the sorting is the whole finding — a single "strip non-letters" pass would destroy scholarship.

The Ross case is the Dresser case from ROADMAP Phase 9 in miniature: the mark is deictic, and deleting it turns a definition into a sentence pointing at nothing.

26,119 rule characters are a table, and a table read aloud one atom at a time is noise. This is the §2 apparatus class, not a symbol problem — it belongs with the Phase 10 pass that withdrew 49 divisions, and it was missed there because that detector reasoned about WORDS. Worth measuring which divisions carry it before deciding: if it is confined to front matter the disposition is obvious, and if it is threaded through the reading it is not.

The Rilke letters are the serious one, and they are not cosmetic. All 14 sit inside German words a reader is currently served broken:

Dr?ckende      persönliche → pers?nliche      lebensf?hig
da?            öffentlich  → ?ffentlich       ?bereinkommen
gel?st         k?nnen      → bed?rfen         versch?tteten

Drückende, persönliche, lebensfähig, daß, öffentlich, Übereinkommen, gelöst, können, bedürfen, verschütteten. Every one is recoverable by hand from context — this is an editorial repair, not a loss.

It is the third payload-damage site recorded. Sextus carries stripped Greek (imperturbability, . We), the Emerald Tablet carries this emoji, and the Rilke letters carry these. The standing note that correcting a checksummed payload "is an editorial act needing a ruling, not a fix" applies to all three, and they should be ruled on together rather than one at a time.

It is also the same family as ROADMAP 12.2, where the Chain tokenizer shreds accented Latin — non-ASCII Latin failing to survive a pipeline, once on the way in and once on the way out.

source-hygiene.test.js sweeps every source file for control characters, and was written after a literal U+0008 hid inside a regex for two sessions. It does exactly what it says, and a printable symbol is invisible to it.

The guard that would catch the next one is an inventory, not a rule. An allowlist of "permitted symbols" would need a judgement per codepoint and would drift toward permitting everything — the failure already named in ROADMAP 12.4, where a test that recovers almost everything by looking almost everywhere has stopped asking a question. Instead commit the measured inventory above and assert the corpus still matches it, the same shape as division-index.json: a new symbol in a future ingest fails loudly, and no judgement about beauty is encoded anywhere.

Reported from reading: "Chapter II of A Hundred and Seventy Chinese Poems is no more than a table of contents, which starts with the text 'Chapter II' but contains no text." Exactly right, and it is three divisions, not one.

a-hundred-and-seventy-chinese-poems serves 17 divisions. The first four are front matter, and three of those are contents blocks wearing chapter names:

CHAPTER II:

      Satire on Paying Calls in August         57
      On the Death of his Father               58
      The Campaign against Wu                  59

Two defects, and the second is the one that hides the first.

1. 282 words of contents served as three readings. Small, and wholly removable — nothing is lost, because none of it was a reading.
2. Every one of those names is served twice. Chapter I, Chapter II and Chapter IV each appear again later carrying the actual poems. This is the Shahnama keying shape — 462 sections resolving to 249 distinct names — where a duplicate name is the visible symptom of a division that should not exist. Note also that no contents block is named Chapter III: the numbering comes from the page being listed, not from what the division holds.

A division at least five lines long, where half or more of its non-empty lines end in a page number after run-out spacing:

/\S\s{2,}\d{1,4}(\s*[-–]\s*\d{1,4})?\s*$/

Swept across all 91 works: 5 divisions, 5,732 words. It is right twice and wrong once, and the wrong one is the largest.

Stated plainly because the ratio matters more than the hits: 4,000 of the 5,732 words this detector proposes are a translator's commentary. Run at this threshold and applied without review it would delete the most valuable text it found. It is a reporting instrument, not a trimming one — the same conclusion §2b reached about running heads, and the same one the imprint detector reached after being wrong three times.

Disposition. The three Chinese-poems divisions are the Phase 10 class A case and are safe to withdraw on the evidence above. The other two are flagged for a human and nothing else. Removing text remains a decision, not a pass.

Status: observed and specified; not yet applied. This finding came from a real Workshop composition rather than a corpus detector. That origin matters: the media score was correct enough to expose both a runtime boundary defect and an archive defect in the same thirty lines.

The first division is named Chapter 1, but its startAnchor is Front matter. Its content begins:

ANNA KARENINA
by Leo Tolstoy
Translated by Constance Garnett

Contents

PART ONE
PART TWO
PART THREE
PART FOUR
PART FIVE
PART SIX
PART SEVEN
PART EIGHT

PART ONE
Chapter 1

Happy families are all alike …

The first PART ONE through PART EIGHT run is navigation apparatus. The second PART ONE and Chapter 1 are genuine structural headings and must be preserved. The clean reading therefore begins at the last PART ONE before Chapter 1, not at the first occurrence of either string.

This is not the §2f shape. It is not a contents page mistakenly served as a whole division; it is a contents preamble fused into the first genuine prose division. A division-level contents detector can therefore pass while the reader still receives apparatus.

Adjacent visual clips were assigned across this material. Before the runtime correction, phrase mode could build one atom across several media endpoints: the first field remained active through PART EIGHT, the intervening visual was never authoritative, and the Chapter 1 visual began only at the next prose phrase. That defect is now guarded in source-span.test.js: score endpoints become hard atomization cuts, the phrase floor cannot merge across them, and a single atom cannot compile to two clips on the same track.

That runtime fix does not cleanse Anna Karenina. Conversely, removing the preamble must never be treated as the runtime fix. The standing rule is:

Archive cleansing improves what the reader receives; Session compilation remains correct even when a dirty source reaches it.

Report the first served division when all of the following are true:

1. it is named as a prose division but begins from a front-matter anchor;
2. a literal contents marker occurs before its first genuine prose;
3. a run of top-level division labels occurs before a repeated first label;
4. the repeated label is immediately followed by the division's own heading;
5. prose follows that second structural opening.

The candidate removal is the bounded prefix before the repeated structural opening. The report must include both occurrences and the proposed retained boundary. It must not mutate automatically: an epigraph, dedication, dramatis personae, translator's preface, or a legitimate Part heading can have similar position without being disposable.

1. Compare the proposed boundary with the pinned Constance Garnett artifact.
2. Prove that PART ONE / Chapter 1 survives byte-for-byte after the trim.
3. Add this exact preamble to the known-bad fixture corpus.
4. Sweep the same report across every first division; read every hit.
5. Record the applied span, checksum change, basis, and date in the work's cleanse log and acquisition dossier.

The likely disposition is a bounded front-matter trim, but removal remains a cleansing decision until this verification is complete.

The guard §1b asked for, and the successor identity.test.js names inside its own exemption list. That list had grown to eighteen works — legitimate books whose opening pages do not name themselves — which is the shape ROADMAP 12.4 warns about: a test that recovers almost everything by looking almost everywhere has stopped asking a question. This one carries no exemptions.

The invariant: no division may be devoid of the vocabulary its own work is dense in.

Every division of every English book is dense in said and great, and so were the war volumes. What separates a work is vocabulary common inside it and rare across the rest of the shelf — Pandava, Yudhishthira, Parva. So a work's characteristic terms are its most frequent tokens that appear in no more than 15% of the other ninety works, capped at 120 terms. A division's score is how many of those it uses.

Below 5,000 the instrument refuses Epictetus, Boethius, Kwaidan and Leaves of Grass for being aphoristic — chapters that genuinely do not use their book's distinctive nouns. At 5,000 the corpus is clean with no exemptions, and the defect the guard exists to catch was 115,312 words with zero hits: a twenty-three-fold margin.

Contents pages and printer's apparatus are a different class at a smaller scale, with their own detectors (§2e, §2f). This one is for a whole division that is a different book.

Measuring needs all 16.3M words tokenised. identity.test.js already pays 22 seconds to load the payloads once, and a second full pass would put the unit suite past four minutes for a corpus that changes on acquisition days only. So it is derived offline, committed, and asserted — the division-index.json shape, for the division-index.json reason. The test costs 1.2 seconds.

Eight of ninety-one works cannot be checked by this measure and are named in the artifact rather than passed quietly: kandinsky-spiritual-in-art, literary-essays-emerson, literary-poems-blake, literary-poems-dickinson, paradise-lost, sacred-emerald-tablet, sacred-rumi, sacred-tao-te-ching. Their vocabulary is not distinctive against the shelf, so there is nothing for the measure to hold. A guard that quietly passed them would be claiming a coverage it does not have.

The coverage assertion. A work added to the shelf without re-running the audit fails here, and the failure names the command. That is the acquisitions cannot bypass the gate requirement, and it is why the artifact lists every work rather than only the offending ones.

Mateo's tactic, and it works on the text he had in mind. A division with a very low word count compared to its siblings. Measured as: under a fifth of its own work's median division, and under 400 words absolute.

It found the Chinese poems contents-as-chapters independently of §2f, which is the useful part — two instruments built from different signals agreeing on the same four divisions:

a-hundred-and-seventy-chinese-poems [0]  212w vs median 2248  (0.09)
a-hundred-and-seventy-chinese-poems [1]  106w vs median 2248  (0.05)
a-hundred-and-seventy-chinese-poems [2]   81w vs median 2248  (0.04)
a-hundred-and-seventy-chinese-poems [3]   95w vs median 2248  (0.04)

Report-only, and it must stay that way. 123 runts across 88 works, and the list plainly mixes defects with genuine short chapters — Moby-Dick's The Lamp. (242w) and Midnight Aloft.—Thunder and Lightning. (50w) are Melville writing short chapters on purpose. This is rung 2 doing exactly what §3 says rung 2 does: producing suspicions, not verdicts.

New suspicions worth a human, beyond those already recorded:

• don-quixote [1–5] — 85w, 250w, 49w, 43w, 106w against a 2,512 median, with names that read as truncated headings: OF THE INTERVIEW THE CURATE AND, OF WHAT BEFELL DON.
• faust [0–2], madame-bovary [0], metamorphoses [0], a-doll-s-house [0] — front matter at the head of the work.

Checked directly. Anna Karenina's first division is 982 words against a 1,428-word median — a ratio of 0.69, entirely unremarkable, and divisions 2, 3 and 4 open cleanly at Chapter 2, Chapter 3, Chapter 4. Only the first carries apparatus, and only about forty words of it, fused ahead of genuine Tolstoy.

That is what §2g already says: a contents preamble welded onto the front of a real prose division. No division-level size or content measure can see it, because the division is the right size and is mostly the right book. It needs a head-of-division trim, which is the §2b shape and Phase 10's class B — trim the opening, never withdraw the division.

Two defects, two instruments, and the distinction is worth holding: §3b asks whether a division is the wrong book; §3c asks whether it is too small to be a reading; §2g is neither, and needs the opening trimmed.

Withheld works keep their payload files, so an early run reported King Lear's Cambridge variorum apparatus as a finding — a text withdrawn for precisely that, in 2026-07. The audit now reads the WITHHELD map out of index.js rather than restating it, and skips those three.

Mateo's ruling: every defect we find must be logged as something that can be looked for again. Don Quixote served bookcover.jpg, spine.jpg (152K) and Full Size as if they were Cervantes — and the moment that is written down as a pattern rather than repaired as an instance, it becomes a question the whole shelf can be asked, today and at every future acquisition.

src/content/archive/defect-signatures.js holds them; scripts/audit-defect-signatures.mjs asks all 88 works all of them.

Provenance is required. Each entry records the work that taught it and the date. A signature without provenance is a guess; with it, a later reader can open the thing itself and judge whether the pattern still describes it.

Disposition is not severity. withdraw — the division was never a reading. trim — a genuine reading carries something in a line. review — evidence for a human, and nothing may be cut on it alone. Nothing in the module applies itself.

Don Quixote is far worse than the five head divisions suggested. Roughly 990 lines of image furniture run through the whole work, not a preamble: p003.jpg (307K) and Full Size recur for hundreds of plates. The §3c runt measure saw only the short divisions at the front; this saw the shape of it.

The loop closed on its own first output. The sweep printed a Rilke line as evidence for replacement-character — and that line also read Zwar fühlen viele junge Menschen. Undecoded HTML entities became the eleventh signature within the minute, and found 89 lines. So those letters carry two encoding failures, not one: fourteen characters destroyed outright and eighty-nine lines never decoded.

A signature at zero is not a signature to delete. variorum-sigla finds nothing, because the three works it describes were withdrawn in July. It is kept, and asserted at zero, precisely so the next acquisition cannot repeat Hamlet unseen. A guard retained only while it fires is not a guard.

Rung 3 — a model reading — belongs on the neighbourhoods these produce, not on whole books. The registry is the targeting system, and it is what makes the cleansing recursive rather than repetitive.

Don Quixote: 626 lines of image furniture, cut line by line. Not a preamble — it ran through the whole work, p003.jpg (307K) and Full Size recurring for hundreds of plates across 145 sections. Every one of the 23 distinct line shapes was listed and read before the cut, and not one was prose. No division was withdrawn: apart from these lines the work is Cervantes throughout. 425,995 words remain and all three signatures now report zero.

Le Morte d'Arthur: three heading lists withdrawn, 6,140 words. Sections that run chapter heading to chapter heading with no prose between them.

The discriminator was checked against its own neighbour rather than trusted:

A contents list runs heading to heading; a real chapter has a heading and then Malory. One measure separates them, and the neighbour proves it.

Both Front matter sections were left alone, deliberately. They carry contents too, but they open with Le Morte D'Arthur / King Arthur and of his Noble Knights… — the title block, and the only place the work names itself in its opening pages. Karenina taught that on the same day: cutting to the first real heading removed exactly that and identity.test.js failed at once. They want the Karenina treatment — cut the navigation run, keep the title — and they get their own pass rather than a hurried one here.

Everything above cleanses the TEXT. The four findings below are about the SCHEME: what a work's divisions are called, where they start, and whether they are divisions at all. That used to be internal bookkeeping. It is not any more.

The Scriptorium can now compose a reading shorter than the shortest whole work by naming a division, or a division's opening (src/core/library-extent.js). A curator names one by number; the reader is told which one by name. So a division named from a bibliography, or a division that is a scanner's header, is no longer a blemish inside a payload — it is what a reader is handed and what the interface calls it.

Five works have lost their line breaks. A poem's lines were joined with spaces and only the stanza breaks survived, so the verse arrives as run-on prose. Blake's quatrains reach the reader as one 117-character line.

The separation is total. Verse either kept its lineation or lost all of it:

Nothing sits between 37 and 99. This is not a spectrum of typesetting quality; it is two groups, and five works are in the wrong one.

The damage does not stop at the line. Headings are found by their shape on a line of their own, so a work whose lines were joined has no headings left to find:

• sacred-rumi — 14,110 words in 35 lines. No scheme was detectable, so RISE cut it itself: reason: "measured", four divisions called Reading. The one work in the library where the divider had to invent, and the reason is upstream of the divider.
• literary-poems-blake — 5,539 words, divided: false, one division. Fifty heading lines stand inside it — its poems, and the book's own two section titles — and not one of them can be named. Its table of contents went the same way: nineteen Innocence titles run together on one line.
• kabir-songs — 11,515 words, divided: false, one division.

So a curator cannot ask for The Tyger. It can ask for Blake whole, or for the opening of the single division that is the entire book — which begins, as the example below shows, with a picture caption and a title page.

And Blake is the work that defines the library's short-length cliff: at 5,539 words it is the second-shortest thing on the shelf, the first being the Emerald Tablet at 237. The one book most needed for a short reading is the one whose poems are locked inside an undivided section.

Our payload is the Astolat Press edition of 1901 (London: R. Brimley Johnson) by way of Gutenberg. What it serves:

THE TIGER

Tiger, tiger, burning bright In the forests of the night, What immortal hand
or eye Could frame thy fearful symmetry?

In what distant deeps or skies Burnt the fire of thine eyes? On what wings
dare he aspire? What the hand dare seize the fire?

Blake engraved it in 1794 in Songs of Experience. The same eight lines:

Tyger Tyger, burning bright,
In the forests of the night;
What immortal hand or eye,
Could frame thy fearful symmetry?

In what distant deeps or skies.
Burnt the fire of thine eyes?
On what wings dare he aspire?
What the hand, dare sieze the fire?

Ten separate corruptions, of two different kinds:

The two kinds do not have the same remedy, and that difference is the point of this section.

1–3 are OURS. The 1901 book had line breaks, a picture rather than a caption of one, and a table of contents. The ingest lost them. These are repairable from the source we already used and nobody has to decide anything.

4–10 are the EDITION'S. R. Brimley Johnson modernised Blake, as Victorian editors did. Repairing those means acquiring a different edition, not fixing this one — and until that is done the shelf serves a modernised Blake under Blake's name with nothing saying so.

& for and, sieze, water'd and the terminal ! are Blake's engraved readings and are not errors to be tidied. The reference text above is set down as the TARGET, from the received transcription of the plates. It is not an acquisition. A poem typed from memory is exactly the substitution this document exists to prevent. Ingesting it means sourcing an edition that prints Blake's own text — an Erdman, or a Blake Archive transcription — with a provenance record like every other work, never a hand-keyed file.

• Lineation (1) — repairable, not yet specified. The five works want a re-ingest from source with line breaks preserved. Whitman, Blake, Dickinson, Rumi and Kabir are the list. Rumi and Blake regain a division scheme by it, which is the larger prize.
• Picture captions (2) — trim. [Picture: …] is the transcriber-image-file class in a different costume: the caption of a plate the payload does not carry. No signature holds this shape yet.
• Modernised edition (4–10) — review, and declare. Nothing may be edited into a payload to make it look older. Either acquire an edition of Blake's text, or say on the card which edition this is.

sacred-corpus-hermeticum has 244 divisions named from a works-cited list.

Volume 1 — D.  J,  L.  =  Mead  (G.  E.  S.),  Did  Jesus  Live  100  B.C.?
F. F. F. = Mead (G. K. S.), Fragments of a Faith Forgotten.  Some Short
Sketches among the Gnostics, mainly of the first two centuries

The scheme is counting bibliography entries rather than tractates. Found while sending division labels to a curator, where a name is the thing being chosen by — a division named for a citation cannot be chosen, and the work has 244 of them.

Recorded as bibliography-as-division-title in defect-signatures.js. The sweep confirms it: 10 lines, one work, and nothing else in 88 works.

Disposition: review. The scheme is wrong, not the text, and re-dividing a Hermetic corpus is an editorial act. Nothing is cut on a signature alone. It is excluded from the curator's labels meanwhile — not because it is long, but because its names are not names.

Thirty-three of eighty-eight works open on something that is not the work. A Gutenberg header, a title page, a contents list — sitting at division one, which is where anything naming a division blindly will land.

extended-bhagavad-gita-full   division 1 = "Front matter"          body from 2
the-kalevala                  division 1 = "Front matter (1/3)"    body from 4
le-morte-darthur              division 1 = "Front matter (1/2)"    body from 3
beowulf, faust, metamorphoses, phaedra, a-doll-s-house, …          body from 2

The divider already labels this matter Front matter — it has always known. Nothing read the label.

division-index.json now carries bodyFrom: the first division that is the work itself, present only when something precedes it. It reaches the curator through divisions.bodyFrom, and the prompt forbids naming a division below it.

The predicate (isFrontMatterLabel, in divisions.js beside where the labels are written) reads only the label this codebase generates. An author's own Preface & Introduction — Ross, A Theory of Pure Design — is a division of the book and a reader may well want it. Detecting front matter by meaning rather than by our own mark would take that decision away from them.

§3d made cleansing recursive by writing each defect down as a pattern the whole shelf can be asked about. A shape defect does not fit in it.

DEFECT_SIGNATURES tests one line at a time, which suits an image filename or a boilerplate line. Lost lineation (§2h) is not a substring — it is a property of a work's line-length distribution. The obvious regex was tried: a long line with four or more mid-line capitalised starts. Measured against the corpus it fires on 35 works, and its worst offenders are prose:

 97%  sacred-rumi                  ← the real thing
 69%  extended-bhagavad-gita-full
 57%  okakura-book-of-tea
 32%  literary-walden              ← Thoreau, correctly typeset
 20%  romance-of-the-three-kingdoms

It was not added. A signature that cries wolf on Walden makes the registry worth less than no registry, because the next person reads a report full of noise and stops reading reports.

The measurement that does separate them is in §2h and takes two numbers per work — mean line length, and the share of lines over 120 characters. It wants a work-level audit script beside the line-level one, taking a declared list of works whose form is verse. Until that exists this section IS the record: the recipe and the five names, so the question can be asked again.

The registry's shape is a finding in itself. Everything in it is a defect you can point at on one line. Everything absent from it is a defect you can only see by standing back, and this is the first of those we have written down.

The worst defect found in the whole campaign, and no detector could ever have found it. It is not garbage in the text. It is text that is gone.

An ingest finds a heading by its shape on a line of its own, and the heading line is then REMOVED from the content — correctly, since a division's title is not its first sentence. Three works were given a pattern with no literal anchor and a case-insensitive flag:

literary-walden           /^[A-Z][A-Z ,'-]{4,55}$/i
literary-poems-blake      /^[A-Z][A-Z ’,'-]{2,55}$/i
literary-leaves-of-grass  /^(?:BOOK [IVXLCDM]+|[A-Z][A-Z ,’'-]{5,65})$/i

i makes [A-Z] match lowercase. So ^[A-Z][A-Z ,'-]{4,55}$ matches any line of five to fifty-six letters, spaces, commas, apostrophes or hyphens — which is to say, most lines of ordinary prose and nearly every line of verse. Each one matched became a division boundary, and was deleted from the body.

Whitman has lost 9.3% of Leaves of Grass — some nine hundred lines. Blake escaped only because its pattern matched so much that no scheme survived validation, so nothing was consumed and the work arrived undivided instead.

Of Walden's 48 division names, seventeen are real chapters. Thirty-one are Thoreau's own prose, and each one is a line no longer in the book:

  "Inde genus durum sumus, experiensque laborum,"
  "The wind that blows"
  "Two second-hand windows"
  "expeditions, using new passages and all improvements in"

Walden quotes a poem about Baker Farm. Project Gutenberg #205, our own cached artifact, line 5798:

     “Thy entry is a pleasant field,
     Which some mossy fruit trees yield
     Partly to a ruddy brook,
     By gliding musquash undertook,
     And mercurial trout,
     Darting about.”

What RISE serves:

     “Thy entry is a pleasant field, Which some mossy fruit trees yield
     Partly to a ruddy brook, By gliding musquash undertook,

     Darting about.”

And mercurial trout, matched the pattern. The word mercurial does not appear anywhere in our Walden.

Every other defect in this document is something PRESENT that should not be: a filename, a running head, a citation, a replacement character. A signature can be written for each because there is something to match.

A missing line has no shape. There is nothing to detect. The text reads perfectly — grammatical, in Thoreau's voice, continuous — and it is not what he wrote. No sweep over our own payloads could ever find it, however clever, and that is not a gap to be closed by a better sweep. It is the boundary of what sweeping can do.

It was found by comparing our text against another edition, one word at a time, and it was found in the first three thousand characters compared.

This is the whole argument for ARCHIVE-CANON-SPEC, and it was found the day after that document was adopted, in a work already in the canon.

• literary-walden — re-source before certification. It is canonical and it is missing 303 words. Certification exists for exactly this.
• literary-leaves-of-grass — withheld, and now with a measured reason. 11,359 words.
• The patterns themselves — fix at the ingest, not the payload. An unanchored capital class with i is the bug; the three works are the symptom. A heading pattern must be anchored on a literal the work actually uses (CHAPTER, BOOK, CANTO), or on a titled scheme read from section names.
• A word-count reconciliation belongs in the ingest. Source words in, payload words out, and a difference beyond the known front-matter trim is a failed ingest rather than a finished one. That check costs nothing and would have caught all three on the day they were made.

Class: a structured heading flattened to a string, and a hierarchy dropped at a file boundary. Both are the ingest reading less than the edition wrote. Neither is visible to word reconciliation, because both lose STRUCTURE while every word still arrives.

Found by a reader looking at the shelf, which is the third time that has been how a defect of this class surfaced.

A Standard Ebooks heading marks its own parts:

<hgroup>
  <h4 epub:type="z3998:ordinal z3998:roman">I</h4>
  <p  epub:type="title">Fyodor Pavlovitch Karamazov</p>
</hgroup>

textContent returns I Fyodor Pavlovitch Karamazov, which is three declared facts pressed into one string with the label — printed by the stylesheet from the section's own epub:type — silently absent. On the shelf that read as a numeral that restarted twelve times and explained nothing.

The parts are read separately now: se:label or the section's unit word, then z3998:ordinal, then epub:type="title", composed as Chapter I: Fyodor Pavlovitch Karamazov. Reconciliation still weighs the heading AS WRITTEN, so a word a stylesheet prints is never counted as one we imported.

Dostoevsky's Books, Eliot's Books and Joyce's Parts are each filed as their own spine document, so a chapter has NO ancestor to read within its own file. The importer only composed ancestry it could see in one document, and so dropped the level entirely — ninety-six chapters, and nothing to say which Book.

The spine is the edition's own statement of what contains what. A container file now opens a scope that governs every reading after it, and closes every scope at its own depth or deeper: Part I holds Books I–III, and Part II must close Book III before it closes Part I.

Composition happens after numbering, not before. Joyce titles none of his episodes; prefixing first left three readings called "Part I" where eighteen belonged.

Not a line-level defect, so the registry cannot hold it (§3g). It is an ingest-time invariant instead, and the ingest now asserts all three:

All fifteen canon works were re-ingested against these. Zero duplicates, zero bare names, zero words lost.

Published from docs/specs/ARCHIVE-CLEANSING-SPEC.md. Edits made here are overwritten on the next push to main.