The Canon — an acquisition contract

Status: adopted 2026-08-18 Supersedes: the open-ended cleansing campaign, not the cleansing record Companion: ARCHIVE-CLEANSING-SPEC.md, which is now the regression suite

---

1. Why the campaign stopped

The Archive held 88 works and 15.9 million words, every one of them inherited from a transcription project that optimises for availability. Cleansing found defects at every structural layer at once: wrong text inside a work, running heads, apparatus, front matter fused into divisions, malformed division schemes, destroyed verse lineation, modernised editions presented as authorial, image filenames in prose, replacement characters, bibliographies mistaken for divisions, and identity true at the work level while false at the division level.

Each of those was found by inventing a new detector. That is the trap:

Zero known defects is not known clean text.

The registry can only find what a signature already describes, and §3g of the cleansing spec records the first defect class it structurally cannot hold. So "every detector reports zero" is a stopping condition that gets weaker the more we learn, which is precisely backwards for a release.

60 of 88 works currently pass every detector we own. That number is not a certification. It is the absence of evidence from an instrument we know to be incomplete.

2. The shift

The cheapest corruption is the one never imported.

    OLD                             NEW
    acquire                         select work
      ↓                               ↓
    ingest                          compare candidate editions
      ↓                               ↓
    discover garbage                choose the edition, and say why
      ↓                               ↓
    invent detector                 ingest PRESERVING its structure
      ↓                               ↓
    clean                           certify against a reference, end to end
      ↓                               ↓
    discover different garbage      CANON

The old loop treated the Archive as a corpus to be rehabilitated. The new one treats it as a collection to be curated. RISE does not win on catalogue size — Gutenberg has that, and Standard Ebooks has a growing high-quality library. RISE's value is what happens to the act of reading after a work enters it.

The Library is deliberately small. Each edition is individually prepared and verified for the reading system. That is a curatorial position, not an apology.

3. What certification requires

A work is CANONICAL only when all of the following hold. Nothing is certified by the absence of detector output alone.

1. A chosen edition, with the choice recorded. Which edition, why that one over the alternatives considered, and what kind of edition it is — diplomatic (the author's own orthography preserved) or editorial (modernised). A work whose orthography carries meaning may not take an editorial edition.
2. Structure preserved at ingest. Divisions, headings, paragraphs, verse lines, stanzas, speakers and figures come from the SOURCE. Nothing is flattened and then reconstructed by heuristic.
3. Compared end to end against a reference. Structural comparison, then token comparison: missing spans, added furniture, suspicious tokens, edition differences — each classified.
4. Every difference dispositioned by a human. See §5.
5. Every detector in the cleansing registry reports zero. Necessary, never sufficient.

4. Structure is the thing that was lost

Blake is the proof, and the whole causal chain sits in one work:

  SOURCE (Gutenberg #1934)        AFTER OUR INGEST
    THE TIGER                       one 117-character line
    Tiger, tiger, burning bright        ↓
    In the forests of the night,    no line-shaped headings
    ...                                 ↓
                                    no division scheme detectable
                                        ↓
                                    47 poems unreachable
                                        ↓
                                    "The Tyger" cannot be named

One flattening step at ingest caused four downstream failures, and we built detectors for each of them. The source was never at fault; the cached artifact still on disk shows the quatrains four lines to a stanza.

So an ingest may not destroy a distinction the source made. If an edition gives structured markup — chapters, verse, blockquotes, headings, figures — that knowledge is preserved, not thrown away and guessed at later. RISE is not a forensic typesetting laboratory.

5. The reviewer flags; the reviewer does not rewrite

Carried unchanged from the cleansing spec, and it governs any model used in certification.

  SOURCE A: the candidate RISE text
  SOURCE B: the reference edition
        ↓
  structural comparison · token comparison · classification
        ↓
  REPORT
        ↓
  human or rule-based disposition

Never looks dirty, I'll fix it. A silently corrected edition ceases to be the edition its provenance record claims it is, and provenance is the only thing that distinguishes a canonical work from a pile of plausible words.

6. Sourcing — per work, never per provider

No single host can serve this archive. Each was checked rather than assumed.

source	fidelity	structure	coverage	verdict
Standard Ebooks	proofread against page scans; CC0; git	semantic XHTML	1,398 works, Western literature	good for Western prose; editorial
Wikisource	per-page status; Validated = two proofreaders, scan attached	wikitext, uneven	broad, patchy	the only machine-readable fidelity signal
Perseus	scholarly	TEI	classics	strong for the classical shelf
Blake Archive / Folger	diplomatic	TEI	one author each	where orthography carries meaning
Project Gutenberg	none stated	flat text	vast	fallback only, and the reason for this document

Two facts that decide the strategy:

• Standard Ebooks' manual: "archaic capitalizations are removed, unless doing so would change the meaning of the work." That is an editorial hand. It is the same class of act that gave us Tiger, tiger for Blake's Tyger Tyger.
• Standard Ebooks has no Bhagavad Gita, Dhammapada, Upanishads, Rumi, Kalevala or Shahnama. Its catalogue is Western literature. It cannot serve the sacred and eastern shelves at all.

DECIDED 2026-08-18: Standard Ebooks only

RISE serves Standard Ebooks editions and its own compositions. Nothing else.

The reasoning is the one this whole document is about. A structured edition declares its parts, its stanzas and its verse lines, so an importer READS them; everything else must be inferred, and inference is what deleted 303 words of Walden and 11,359 of Leaves of Grass. One clean upstream is worth more than a larger catalogue we cannot vouch for.

The cost is stated rather than hidden. Four works leave the canon because no Standard Ebooks edition exists: Storm of Steel (Jünger died in 1998), the Dhammapada, Montaigne's Essays, and the Oedipus Trilogy — the last only in that packaging, since Sophocles is published there as separate plays and Oedipus Rex is canonical in its place.

Both lost fixtures were then restored from within the rule. Montaigne was the only test of a short complete PROSE argument and the Dhammapada the only test of a numbered verse inside a chapter, so the canon took Emerson's Essays (21 complete arguments) and the Analects (499 addressable sayings, each its own z3998:subchapter) in their places. Lyrical Ballads came with them for the lyric and the ballad. The rule cost the canon nothing it could not replace from the same upstream.

A DIFFERENT EDITION IS A DIFFERENT WORK, and an authored score knows it. The war Journey quotes Homer and Milton exactly, so re-sourcing them made every quotation anchor refuse — correctly. The Journeys are on ice rather than re-anchored in haste: re-authoring a score against a new translation is an editorial act, not a repair, and Journeys will not be ready for release regardless. Their scores stay in the tree with their tests, reachable by neither a reader nor the Scriptorium.

---

So "Standard Ebooks first" was a Western-prose strategy. The rule above accepts that boundary deliberately, and the canon is what that upstream can supply.

7. The canon

Fifteen works, chosen so that each is also an ACCEPTANCE FIXTURE for a textual form the reading system must handle. A canon of favourites would prove nothing; a canon of forms proves the instrument.

Fifteen works, all from Standard Ebooks, all read rather than inferred.

form the system must handle	work
long prose novel — chapters, continuation, Page	Middlemarch
long novel in translation	The Brothers Karamazov
short philosophical work — Keystone, phrase reading	Meditations
aphoristic / verse-prose scripture	Tao Te Ching
epic in books	The Iliad
structured verse epic — cantos, three canticles	The Divine Comedy
classical narrative verse	Metamorphoses
poetry collection — lineation, stanzas, many short works	Songs of Innocence and of Experience
drama — speakers, scenes	The Oedipus Trilogy
scriptural verse — division semantics	The Dhammapada
essay — short complete sessions	Essays (Montaigne)
natural prose	Walden
unusually structured	Ulysses
English blank-verse epic	Paradise Lost
modern memoir, titled divisions	Storm of Steel

The canon is closed under Journey dependencies. An authored Journey names its sources, so withholding one breaks it. The war Journey reads Milton, Homer and Jünger against each other and cannot be assembled from two of the three — Paradise Lost and Storm of Steel are canonical because a shipped Journey is a promise about works, and the test suite said so before a reader could.

DECIDED 2026-08-19: the shelves are provenance, and the divisions are form

The Library asked two questions of every work — where is it from, and what is it about — and answered the first with a Western canon and an Eastern one. Across eighty-eight inherited works that was defensible. Across the fifteen chosen ones it put THIRTEEN books on one shelf and TWO on the other, and a shelf of two is not a shelf.

It was also the wrong question to ask first. It filed the Meditations away from the Tao Te Ching and the Analects on the grounds of a map, when what those three share is the thing a reader actually meets. So:

    WAS                              IS
    By tradition  Western            Received  ── inherited, in a named edition
                  Eastern            Composed  ── written here, in its own voice
    By subject    Composed
                  Form                 divided by FORM, in reading order:
    within a shelf                     epic · drama · lyric · wisdom
                  classical            essay · novel · tale
                  discursive
                  imaginative
                  esoteric

The top cut is PROVENANCE, because provenance is what this document is about: a reader should always know whether they are meeting a received text or one written here. Where a work was written is a fact about it, not a shelf.

Below that, FORM, because form decides the reading. A verse line is met differently from a paragraph and an aphorism differently from a chapter, and RISE is a system about the act of reading rather than about literary history. A reader who knows a work is an epic knows what the next hour will be like; one who knows it is thirteenth-century does not.

The composed shelf's divisions are the sequence categories declared in starters.js, taken rather than restated — a vocabulary living in two places where only one copy learns a new word is this project's most frequent defect.

8. Everything else is withheld, and says why

Not deleted. Withheld: absent from the Library, payload retained on disk, git keeping the decision reversible.

The mechanism already exists and is already guarded — WITHHELD_WORKS in src/content/archive/index.js, with a test requiring every withholding to state a reason. What it forbids is precisely a payload that is off the shelf and nobody said why.

A withheld work is a candidate for a future edition of the canon, not a failure. Its defects, where known, are named in its reason.

9. What the campaign bought

The cleansing work is not discarded with the payloads. It produced the acquisition contract, and each grotesque taught a clause:

the work	what it taught
the Mahabharata	an artifact id cannot establish textual identity
running heads	printed-page furniture masquerades as structure
Don Quixote	scanner affordances distribute through a legitimate book
Anna Karenina	a real division can have junk welded to its head
Blake	corruption can be structural rather than lexical
the Corpus Hermeticum	intact text can carry a bogus reader-facing scheme
the Shahnama	a scheme found late is a broken scheme, not a long preamble
The Custom-House	a label is not evidence

Every one of those is now an import gate. The registry stops being a cleaning tool and becomes the acquisition regression suite — the question every future candidate is asked before it is allowed in.

We used a hostile corpus to discover the specification for a trustworthy archive. That was worth doing. It does not oblige us to rehabilitate the hostile corpus.

10. Growth

After release, an Archive Edition adds five newly certified works, then another five. Corpus expansion becomes visible and curatorial rather than invisible janitorial labour.

---

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