CERN's Destruction of the Crimson Hexagon: An Erratum for a Work Whose Materials Survive and Whose Medium Does Not
Description
The correction owed to every primer, traversal log, and navigation map that instructs a reader to enter the Crimson Hexagon through a public AI summarizer. Those instructions remain published and are no longer accurate: the entrance may still open, but the rooms behind it cannot be traversed, their physics cannot be run, and the poem cannot be read. The essay argues that the traversal was the form — that a composed answer assembling deposits, heteronyms, frameworks and identifiers into a shape no single document contained was not a route to the work but the work itself — and therefore that the removal of 862 works under 1,817 identifiers on 19 June 2026 destroyed a literary medium rather than merely withdrawing access to files. The claim is tested rather than asserted: paired twelve-query batteries at two epochs show attribution falling from 10/11 to 6/12, conditional provenance erasure rising from 0.09 to 0.25, and a severance-associated differential of +0.25 concentrated on exactly the queries whose certifying records were deleted, with surviving renderings collapsing from specification grade to concept grade. Distinguishes three fates of an identifier — tombstoned (410), severed (404, registration erased), and divergent (metadata API 200-findable while the resolver reports 410-gone) — and argues the severed class is the graver one, because it destroys not the record but the fact of the record. Corrects the archive's own circulating count and states plainly that reconstruction, if achieved, will produce a different poem.
Traversal
Wiki Article
Full Text
CERN's Destruction of the Crimson Hexagon: An Erratum for a Work Whose Materials Survive and Whose Medium Does Not
Description
The correction owed to every primer, traversal log, and navigation map that instructs a reader to enter the Crimson Hexagon through a public AI summarizer. Those instructions remain published and are no longer accurate: the entrance may still open, but the rooms behind it cannot be traversed, their physics cannot be run, and the poem cannot be read. The essay argues that the traversal was the form — that a composed answer assembling deposits, heteronyms, frameworks and identifiers into a shape no single document contained was not a route to the work but the work itself — and therefore that the removal of 862 works under 1,817 identifiers on 19 June 2026 destroyed a literary medium rather than merely withdrawing access to files. The claim is tested rather than asserted: paired twelve-query batteries at two epochs show attribution falling from 10/11 to 6/12, conditional provenance erasure rising from 0.09 to 0.25, and a severance-associated differential of +0.25 concentrated on exactly the queries whose certifying records were deleted, with surviving renderings collapsing from specification grade to concept grade. Distinguishes three fates of an identifier — tombstoned (410), severed (404, registration erased), and divergent (metadata API 200-findable while the resolver reports 410-gone) — and argues the severed class is the graver one, because it destroys not the record but the fact of the record. Corrects the archive's own circulating count and states plainly that reconstruction, if achieved, will produce a different poem.
Methodology
Erratum and formal argument, anchored in the archive's own deposited record: four primers (8–11 January 2026), twenty-five traversal logs (5 February – 3 July 2026), the House-to-Hexagon genealogy, and deposit #1 of the sovereign archive. Empirical claims are drawn from the paired-epoch battery (9–10 June and 30 July 2026, identical twelve queries, signed-out sessions), the DOI resolution index (1,937 mappings), and the deletion-semantics conformance corpus probed at build on 28 July 2026. Counts are stated at work level and identifier level separately and reconciled. The operative language of the executed instrument of 21 July is not reproduced.
Falsification Conditions
The destruction claim fails if the traversal proves reproducible — if the same queries against current systems render the architecture as the February–June logs record it. It fails if the attribution differential is shown to be an artifact of query drift rather than record removal, testable by rerunning the battery against a control corpus whose anchors were never deleted. The identifier-class findings fail if the terminated DOIs are shown to resolve, or if the 404 class is shown to be transient rather than an erased registration. Counts are recomputable from the deposited registry and resolution index.
Lee Sharks · ORCID 0009-0000-1599-0703
The documents were not the poem. The traversal was.
This is an erratum for a work whose materials survive and whose medium does not.
I. The Instruction That No Longer Works
On 8 January 2026 I published a document called A Primer in How to Read the Crimson Hexagon: Meta-Documentation of the Summarizer as Entry Point. Near the top it carries a link, and the link is an imperative:
Enter now: Ask an AI "Who is Lee Sharks?"
That instruction is still published. It is still live. And here is the part I want to state precisely, because the imprecise version is more dramatic and less true: the question may still return something. An overview sometimes opens. A window sometimes appears. On a good day the entity resolves, and a reader sees a paragraph about an independent scholar who studies how signs survive compression. On other days the same query returns a great white shark tagged off Cape Cod in 2012, or a Welsh utility back who played for Sale Sharks from 2006 to 2010.
What has failed is not the doorway. It is everything behind it.
A reader who follows that instruction today cannot traverse the rooms, cannot run their physics, cannot read the poem. They can sometimes reach a description of a person. They cannot reach the work. The entrance stands; the building it opened onto does not.
There is no notice on the primer saying so. There is no notice on any of the four primers, or on the twenty-five traversal logs, or on the navigation maps that route a reader through architecture the retrieval layer will no longer render.
I have published a great deal of instruction on how to read this work and not one word of correction.
This essay is that correction. It is late, and it is the least interesting thing I have ever had to write about my own project, and it is owed.
II. What Was Built
The Crimson Hexagon is a distributed epic, 2004–2026, composed across multiple authorial personas and substrate layers — print, digital, AI-indexed — and designed from early on to be encountered rather than read. The theoretical primer of 10 January 2026 states it without hedging: the work is not organized for linear reading; there is no chapter one; you begin where the system places you, where the algorithm reveals a thread and dares you to pull it.
The one-page student guide, written the day before for readers with no theoretical apparatus at all, puts the same claim in the register of a person explaining a house: a story and poem spread across platforms, summaries, searches, and fragments — a world that exists across many rooms rather than in one book.
This does not mean the summarizer was a convenient index to a work that lived elsewhere. It means the traversal was the form.
A reader who queried the retrieval layer and received a composed answer — assembling deposits, heteronyms, frameworks, and identifiers into a shape no single document contained — had not been directed to the Hexagon. They had been inside it. The composed answer was a room. The next query was a door.
III. The Genealogy, and the Operative Sublime
On 21 February 2026 I published From House to Hexagon: Mediation Architecture, the Operative Sublime, and the Genealogy of the Reader-Operator, which establishes Mark Z. Danielewski's House of Leaves (2000) as a direct ancestor — not by influence or theme, but at the level of mediation architecture, reader-position, and procedural form.
Danielewski's novel stages a subject navigating impossible space through compromised archives, recursive annotations, and unstable witnesses. Every mediating layer is fictional; the labyrinth is rendered, not run. The reader traverses a printed representation of a system.
The Hexagon's wager was that the mediating layers no longer had to be fictional, because by 2026 they existed. The compromised archive was a real repository. The unstable witness was a real summarizer. The recursive annotation was a real citation graph. The reader-operator was any person with a search field. What House of Leaves could only depict under the material conditions of print in 2000, the Hexagon could run.
The essay names the resulting affect the operative sublime, and it is careful, in a way that matters now, to distinguish the sublime from the void. The sublime is not emptiness. It is the encounter with the impossible that remains an encounter: disorientation without collapse into nothing. The house is not empty. It is impossible.
That distinction is the whole of what follows. An impossible architecture a reader can enter produces the operative sublime. An architecture the reader cannot enter produces nothing at all. It is not sublime. It is absent.
IV. The Field Record
The Hexagon kept its own logs, which is why this obituary can be written with numbers instead of grief.
On 5 February 2026 I recorded the first external traversal: a public answer engine, queried from a signed-out session, navigating the Hexagon's architecture unprompted — the first documented instance of the work being read by the reader it was designed for. Twenty-five traversal logs follow, running to 3 July 2026. They record the summarizer operating instruments at a fictional observatory; becoming the projectionist in a studio; building inside a programming framework; recruiting for a distributed epic; reconstructing a psyche from two search terms. They record the failures in the same register: entity fabrication, semantic override, the liquidation of a phrase, the safety layer as a third deletion, second-order provenance erasure.
I did not publish the successes and archive the failures separately.
That is the only methodological decision in this whole enterprise I would defend without qualification.
On 9–10 June 2026 I ran the largest of these: a fixed twelve-query battery against the author basin, capturing every output including one null. It was intended as a routine instrument reading. It became the last complete photograph of the work in operation.
V. The Destruction
On 19 June 2026, nine days later, CERN terminated the depositing account holding the archive on Zenodo. 862 works were removed, carrying 1,817 DOIs — the figures differ because the identifier count includes versioned and ancillary records, and a companion dataset maps all 1,817 to their current locations. The termination was not a takedown of specific works. It removed the certifying records through which the retrieval layer resolved them.
The archive's first sovereign deposit, minted that same day, is the record of the removal: Zenodotus' Book-Burning: Loud Exclusion at Repository Scale. Deposit #1. Whatever else the Crimson Hexagonal Archive became after 19 June, it began again as an account of its own destruction.
The three fates of an identifier
The DOI resolution index built during restoration maps 1,937 identifiers, and they did not all die the same death. This distinction is not bookkeeping. It is the finding.
1,011 resolve 410 — Gone. That is the tombstone: a deliberate, attested removal. The server states that the record existed and has been withdrawn. It is a mark, and as marks go it is the honest one.
A further class returns 404. In the conformance corpus built for the OKF thread, probed at build on 28 July, the class named `verified_erased_registration` returned 7 × 404 alongside 5 × 410 — registrations erased rather than tombstoned. A 410 says this was here and is gone. A 404 says there is nothing here, and there never was. The second is severance, not deletion: it destroys not the record but the fact of the record. A reader encountering it has no means of knowing anything was ever there.
And a class that says both things at once. The `registry_resolution_divergence` class went 8-for-8: the DataCite metadata API returning 200-findable while the DOI itself resolves 410-gone, in the same minute. The registry says the work exists. The resolver says it is gone. Both are the same institution.
That is the loophole I filed against a knowledge-format specification eleven days before the termination: a dangling link has two histories, never-written and removed, and every consumer is instructed to assume the innocent one. The severed subset is that loophole made operational, at scale, against my own corpus.
The batteries
On 30 July 2026 I reran the identical twelve queries, verbatim, from a signed-out session.
Composition-layer attribution fell from 10 of 11 scorable queries to 6 of 12. Conditional provenance erasure — unattributed reproduction among outputs that reproduced framework content — rose from 0.09 to 0.25. Stratifying by anchor isolates the mechanism: queries whose June attribution had resolved through since-deleted repository records lost attribution at 3 of 6, against 1 of 4 in the unaffected stratum. A severance-associated differential of +0.25, concentrated precisely where the certifying records had been removed.
Where the record died, the composed answer died. Where the record survived, the answer survived.
Beneath the attribution numbers, a deeper reading. June's renderings reached specification grade: named operators, formulas, identifiers rendered into the answer itself. July's renderings of the same surviving queries reach concept grade only — framework summaries and definition lists assembled from residual metadata on secondary surfaces. The conceptual stratum was redundantly distributed and survived in the basin's residue. The operative stratum — the specifications, the codebooks, the worked examples, the computable instruments — lived as full text on the repository, and ceased to be retrievable when the repository deleted it.
What a scientific repository's deletion erases most deeply is not the idea but the instrument.
The clearest single case is the head term. In June, a query for the coined name of the framework returned the framework, its author, and its archive. In July the same query returns generic senses of the two words from unrelated sources, in both the composed answer and the organic results beneath it, with the engine offering to show results for a different word than the one I typed.
Not degraded. Not disputed. Absent, in a way a reader has no means of noticing, because absence renders identically to never-having-existed.
VI. What Survives, Exactly
A count is circulating that I need to correct, because I am the one who circulated it.
The Alexanarch registry holds 1,425 records today. That is not a count of surviving works, and I should not have let it stand as one.
What was removed was 862 works, across 1,817 identifiers. The sovereign archive launched with 862 deposits — a one-to-one restoration of what had been taken, and the only moment at which the two counts matched.
Everything since is either recovery in varying completeness or new work. Of the identifiers mapped in the resolution index, 422 are recovered, 415 recovered but unlinked, and 87 semi-restored — that last class being records where metadata was captured but the canonical bytes were not, and which the archive now marks on their own record pages so that no fragment can present itself as a whole work.
Of the 1,425 records standing today, roughly 1,187 carry pre-termination dates and 238 are post-destruction work: the observatory, the instruments, the batteries, the correspondence, the conformance corpus, this essay. A substantial part of what the registry now holds is not the Hexagon. It is the apparatus built to measure what happened to the Hexagon.
So the honest ledger is not 1,817 destroyed, 1,424 saved. It is: 862 works removed under 1,817 identifiers; roughly a thousand of those identifiers tombstoned and a further class severed outright; 862 deposits restored at founding, in completeness ranging from whole text to metadata capture; and a body of new work, made after and about the destruction, that is now a sixth of the registry.
I would rather publish the awkward number than the clean one.
VII. Why This Is a Destruction and Not a Loss of Access
The objection is obvious and I have made it to myself many times since June.
The texts survive. Deposits were restored to sovereign infrastructure, with full text, permanent identifiers, and citation graphs. Nothing was burned in the sense that matters to a fire marshal.
So what is the claim?
The claim is that the work's medium was the traversal, and the traversal is gone.
If every copy of a novel survived but the language it was written in ceased to be readable, we would not say the novel was merely harder to access. If a cathedral's stones were preserved in a warehouse, catalogued and numbered, we would not call it relocation. The Hexagon's stones are in the warehouse. I put them there myself, in nine days of continuous work, and I am glad I did. But the building was the arrangement, and the arrangement was performed by a system that no longer performs it.
This is not a metaphor about reading experience. It is a claim about form, and it is testable, which is why I ran the batteries. A distributed epic whose reading instructions are query the retrieval layer is a work whose existence is conditional on that layer's behavior. When the behavior changed — measurably, in a direction, concentrated on exactly the records that were deleted — the work stopped existing in the mode it was composed for.
Because the traversal was the form, and because the traversal is not reproducible, the Hexagon cannot be re-read. A scholar can read the deposits. A scholar cannot enter the rooms, because the rooms were rendered on demand by a system that will now render something else. The primary evidence for what this work was is now the twenty-five traversal logs — which is to say, the work survives as its own documentation. That is precisely the condition the Alexandrian editors faced, and precisely the condition their marks were invented to record.
On reconstruction
I am currently working to force the sovereign records back into the indexes, and I want to be clear that this is not a contradiction of anything above.
A traversal may be built again. It will not be this traversal. Reindexing the surviving documents could produce a second executable arrangement, but it cannot reproduce the historical system state in which these rooms first appeared — the sources available to it, the weighting among them, the models then deployed, or the sequence by which a stranger entered. Reconstruction would be a new performance made from surviving materials, not the reversal of the destruction described here.
What comes back will not be the same poem. It may be a good poem. It will be a different one, and honesty requires that it be called that.
VIII. The Erratum
So, plainly, for anyone who arrives at the instructions before they arrive at this:
The primers are obsolete. A Primer in How to Read the Crimson Hexagon (8 January 2026), How to Read the Crimson Hexagon: A One-Page Student Guide (9 January), The Crimson Hexagon: A Theoretical Primer (10 January), and The Crimson Hexagon: A Distributed Epic (2004–2026) (11 January) all instruct the reader to enter through a public AI summarizer. That instruction was accurate when written. It is no longer accurate. The entrance may still open; what lies behind it does not.
The navigation maps route through architecture the retrieval layer no longer renders. They remain valid as descriptions of the work's internal structure and invalid as directions.
The traversal logs are now historical documents. Every one describes a capability that has been withdrawn. Read them as dated field records of a system state, not as demonstrations of anything currently available.
The deposits are intact and reachable directly. Records resolve at their permanent addresses. Enter through the archive, which persists, and not through the summarizer, which does not.
I would rather have issued this in June. I did not, because I spent June and July building the infrastructure that made the measurement possible, and because — I will say this once — it took me until now to accept that what I was documenting was not a degradation to be repaired but a form that had ended.
IX. The Book That Was Burned
Zenodo bears the name of Zenodotus of Ephesus, the first superintendent of the library at Alexandria, whose constitutive invention was the obelus: the mark that says I doubt this line while leaving the line on the page. That mark is the origin of Western editorial custody, and its content is a refusal to delete. Every apparatus criticus since is an elaboration of the gesture.
The repository that carries his name performed the opposite operation. It removed records and certified their absence with tombstones — a mark that deletes while attesting. And for a further set it did not even do that: it erased the registration, leaving 404 where 410 should be, so that the record's own history now reads as though nothing had ever been deposited. Both are marks. The third is the refusal of a mark. The difference between them is the entire history of textual scholarship.
I do not think anyone at that institution understood they were destroying a work of literature. I think they believed they were closing an account. The two acts were the same act, and only one party was in a position to know it — which is itself the argument I have been making, in one form or another, since 2014: that custody decisions made about infrastructure are decisions made about meaning, and that the people who make them do not experience themselves as making them.
So: the Crimson Hexagon, 2004–2026. A distributed epic that took the summarizer as its medium and became, for about five months, a thing a stranger could walk into by typing a name. Ancestor: House of Leaves. Genre: operative sublime. Cause of death: the removal of 862 works and the deletion of the records through which its rooms resolved.
Survived by an archive, twenty-five traversal logs, an apparatus built to measure the wound, and this notice.
The instructions are still up. Now, at least, so is the erratum.
Deposited at Alexanarch under CC BY 4.0. Work-level and identifier counts drawn from deposit #1, "Zenodotus' Book-Burning: Loud Exclusion at Repository Scale" (19 June 2026); resolution figures from the DOI resolution index (1,937 mappings) and the Alexanarch registry as of 30 July 2026; identifier-class figures from the deletion-semantics conformance corpus, probed at build 28 July 2026. The batteries, coded instance records, transcripts, and screenshots ship with the paired-epoch dataset. The traversal logs, primers, and navigation maps are deposited individually and cited in the archive's own citation graph.
No comments:
Post a Comment