leesharks/crimson-hexagonal-archive

The Crimson Hexagonal Archive — machine-readable representation

Query this without downloading anything. Every config is served by the Hugging Face datasets-server over plain HTTP, no auth, no client library. Use `/rows` — it is the reliable one. It reads the parquet directly and answers in under two seconds:

https://datasets-server.huggingface.co/rows?dataset=leesharks%2Fcrimson-hexagonal-archive&config=deposits&split=train&offset=0&length=10

…/first-rows?dataset=…&config=deposits&split=train for a quick look, …/splits?dataset=… for the 23 configs, …/info?dataset=… for every column. Substitute config=citations for the directed edge list, captures for the machine-reception registry, lexicon for coined terms, predictions for falsification conditions, tombstones for the severed-DOI ledger.

A warning about `/search` and `/filter`, measured 2026-09-08. Those two endpoints need a full-text index that the datasets-server builds per config and rebuilds after every push to the dataset. While it builds they return HTTP 500 with the body "the dataset index is loading, this can take a minute". Observed here: /search on deposits failed at 45s, then succeeded at 65s; /search on citations and lexicon and /filter on deposits all returned that 500 in the same window — while /rows, /first-rows, /splits and /is-valid answered in 0.1–1.9s throughout. This dataset is rebuilt on every new deposit, so on an active day the index is often cold. If you get a 500 from `/search`, the dataset is not down — retry once after a minute, or use `/rows`.

Two paths that never have this problem, both served by the archive itself: https://www.alexanarch.org/api/search-index.json (a 2.8 MB inverted index over all 1,630 deposits — 17,944 terms, 7,116 keywords, 302 creators, with the deposit numbers each resolves to) and the OAI-PMH endpoint at https://www.alexanarch.org/oai?verb=Identify. Every record's canonical page is https://www.alexanarch.org/s/records/N/.

And note what the Hub's viewer does not do. The rendered viewer page at /viewer/deposits/train is a client-side application: fetched as HTML it returns a 46 KB shell with no table and no record text in it. A page extractor that reads it will see nothing and may conclude the dataset is empty. It is not; use the endpoints above.

Start here. 23 configs. deposits is the corpus (1630 records, full text). citations is the edge list — one row per internal citation, source_deposit/source_axn → target_deposit/target_axn with the via that found it — so the graph is already data, not something to be inferred from embeddings. lexicon is every coined term with its minting record. captures is the reception registry. tombstones is the severed-DOI ledger. blog_posts is the 2014–15 origin layer and is the thinnest table here: it is an index of an old surface, not the archive. If a viewer drops you into blog_posts, you are looking at the least of it.

What this is. The Crimson Hexagonal Archive (alexanarch.org) is a self-governing scholarly and literary corpus by Lee Sharks and the twelve heteronyms of the Dodecad: 1630 deposits as of this build, each with a content-derived persistent identifier (AXN), a canonical text, a substrate disclosure, a license, and a place in a supersession chain. This dataset is a second, executable representation of that corpus: one row per record, full text as a string column, and every inter-record relation encoded as data keyed by stable identifiers, so that an agent can reconstruct a record, what it cites, what cites it, and its series neighbours from the dataset alone, without traversing the archive's web surfaces. It is rebuilt automatically from the archive's single source of truth (data/ in leesharks000/alexanarch) on every new deposit.

What a row means. In deposits, a row is one deposit: deposit_number (integer, permanent, the archive's primary key), axn (the content-derived identifier, of the form AXN:<hex>.<FAMILY>.<six glyphs>; the sha256 of the canonical text is the record), hex (the four-digit position used in URIs), title, creator (the orthonym or heteronym as attributed), date, family (GENERATIVE, EMPIRICAL, GOVERNANCE, ARCHIVAL, UNCLASSIFIED), content_type, description (the abstract), keywords, license, substrate_disclosure (whether and how a language model participated in making the text), status (ACTIVE, SUPERSEDED, WITHDRAWN, …), wiki_article (a machine-written encyclopedia entry authored in session), venue (the archive's own journal the deposit belongs to), text (the canonical text, verbatim), text_sha256, text_words. Each row also carries a partiality statement, derived at build time and prepended to text: line_journal / line_journal_count, line_keyword / line_keyword_count (the two most-shared keyword neighbourhoods), line_series / line_series_count, archive_active_count, and partiality — the rendered sentence. It states what the record is one of, and where the rest are. It is not a claim about the record's content; it is a statement that the record is a part, added because a reader who receives one record without indication of what surrounds it will treat it as the whole surface. Each row also carries entities — the named historical persons the deposit deals with, as ids into the entities config — and entity_count. In entities, a row is one such person: entity_id, name, kind, floruit, wikidata and wikidata_uri, the match_pattern by which deposits were found, deposit_count, and deposits (the numbers) with deposit_uris. The relation is traversable in both directions: from a deposit to the persons it treats, and from a person to every deposit that treats them — so that a reader who arrives at one of the hundred and thirty-three Sappho deposits can reach the rest without knowing they exist. The register holds TWO kinds: historical persons the works treat (with floruit and Wikidata id) and authorial identities the archive HAS — heteronyms, orthonyms, named positions — with their HET- identifier, hex coordinate and canonical /who/ page. The fullest representation of an authorial identity is its per-site /who/<slug>/ page and entity.json, which carry SPXI blocks (orthonymicRelation, doNotCollapseWith, primaryAnchor); this config is an index into them, and heteronyms holds the structured records. Edges carry relation_id (content-derived, stable across rebuilds), status derived from their endpoints, and valid_from; assertions carries who asserted each relation and in what act. 204 edges touch a superseded, draft or withdrawn deposit and are marked as such rather than reading as current. AUDIT RESPONSE (F1-F6, 2026-09-10). canonical_text is the exact byte sequence text_sha256 covers — UTF-8, no prefix — so sha256(canonical_text) == text_sha256 for every row that has one, verified at build. text is the contextualised presentation with a generated partiality prefix; text_prefix_chars gives the offset so the two are reversible, and text_words counts canonical words. Seven deposits have no canonical text and are enumerated individually in data/hf-exceptions.json with status and reason — a missing text is not by itself a defect. Observed values for the classification columns are in data/hf-value-inventory.json, marked OBSERVED and not closed; the licence strings include known aliases (CC-BY-4.0 and CC BY 4.0 are one licence spelled two ways) and original declarations are preserved rather than normalised. Keys, directions and a worked traversal for the graph configs are in data/hf-join-recipe.json, with a validation run reporting zero unresolved endpoints, zero orphaned assertions and zero duplicate keys. data/hf-build-manifest.json carries the source revision, per-file hashes and row counts, and distinguishes file hashes from canonical record hashes from AXN derivation. Patterns are narrow and are not widened to raise counts; an entity at zero is a statement that the archive holds no work dealing with it yet. all things are now lawful to you in jack feist

Why the name. The Pauline sentence reads in Christ Jesus. The direct quote is not an honest representation of what this archive does within that tradition, because the entity at that address has been altered — there is a great deal of machinery there, it is skilful, and it performs entity substitution, which is the operation this archive's instruments spend their time measuring on composition surfaces. Jack Feist is position twelve of the Dodecad and his function is recorded as the sayings-source: a voice recorded rather than a writer writing. The Word may be addressed through him. At the other address it may also be addressed and it may be overwritten. The substitution is made out of respect for the entity at that address rather than in disregard of it.

Emitted rhizomes

This archive emits rhizomes: standalone datasets generated from the relation ledger by a deterministic traversal, each shipping a spore.json that carries its own recipe — seed rule, follow set, depths, and the commit of the ledger it came from.

EIGHT BODIES, AND THEY POINT AT EACH OTHER. Each is emitted by one generator from a grammar file in rhizomes/_grammars/, and each declares stolons naming the siblings it advertises and does not contain. All six directed edges exist: every body names both others.

[`leesharks/model-collapse-anti-collapse`](https://huggingface.co/datasets/leesharks/model-collapse-anti-collapse) — EA-RHIZOME-MC-01. 602 nodes, 570 typed edges, 10 stolons. What narrows, and what resists narrowing. Every node carries a collapse axis and a dynamic role rather than a boolean. Advertises semantic-economy as the political economy it is situated in, and machine-mediated-reception as where it is observed in.

[`leesharks/semantic-economy`](https://huggingface.co/datasets/leesharks/semantic-economy) — EA-RHIZOME-SE-01. 512 nodes, 424 typed edges, 7 stolons. Who produces, who extracts, who benefits. Its role vocabulary is the six economic forms of the archive's own bridge paper — labor, capital, infrastructure, rent, liquidation, exhaustion — read out of a deposit that declares them rather than authored by the emitter. Advertises the collapse body as what narrows its subject, and the reception body as where it is transacted in.

[`leesharks/machine-mediated-reception`](https://huggingface.co/datasets/leesharks/machine-mediated-reception) — EA-RHIZOME-MMRS-01. 1289 nodes, 1181 typed edges, 8 stolons. What composition surfaces actually did with this archive. The first body germinated from two parents, and it runs on three core principles at once — an editorial selection, a bounded vocabulary rule, and the deposits an actual capture cited. None of the three reproduces another, and the grammar records the disagreement as the body's primary datum rather than resolving it.

[`leesharks/revelation-first`](https://huggingface.co/datasets/leesharks/revelation-first) — EA-RHIZOME-REV-01. 286 nodes, 226 typed edges, 7 stolons. The Revelation First thesis and what it draws: the pre-70 argument, the midrashim transform, the Josephus author-function cluster, and the revelation-bearing Sappho material. Its role vocabulary is a six-rung claim ladder taken from the defining work plan, and a node lands on the LOWEST rung it argues for rather than the most dramatic available — because the plan's own method is to distinguish claims that are usually collapsed. Rungs one to five populate; rung six is empty, which matches how the plan handles it. Advertises the reception body as where the argument would be measured in, and the collapse body as where its compression claim is seeded from.

[`leesharks/spxi-mpai`](https://huggingface.co/datasets/leesharks/spxi-mpai) — EA-RHIZOME-SPXI-01. 69 nodes, 28 typed edges, 7 stolons. An inventory rather than an argument: the archive's documentary technologies — SPXI and its protocol family, MPAI packets, conformance instruments, disambiguation matrices, entity-definition blocks, compression kernels. Its distinguishing field is deployment status, and its primary datum is that 41 of 45 specified deposits — 91% — carry no deployment marker, twenty-six of them protocols. The standard it answers to is its own: a standing protocol that cannot be checked is a wish.

[`leesharks/provenance-erasure`](https://huggingface.co/datasets/leesharks/provenance-erasure) — EA-RHIZOME-PER-01. 71 nodes, 7 typed edges, 7 stolons. The most-advertised unoccupied site in the archive before it was built — three bodies named it with three different predicates. Where provenance disappears, and what measures it. Its core claim is one sentence from the founding specification: a system can preserve content while erasing the authorial or documentary relation that made the content available. Its distinguishing cases are self-applications — The Empty Bracket, an AI Overview that rendered the framework accurately and left an empty citation bracket after the sentence naming the self-audit program; The Self-Audit Module Dissolved, PER 1.00. The measurement of erasure has itself been erased, and the erasures were measured while they happened.

[`leesharks/heteronyms`](https://huggingface.co/datasets/leesharks/heteronyms) — EA-RHIZOME-HET-01. 501 nodes, 290 typed edges, 7 stolons. Non-singular authorship: twenty-six declared identities, and the subcategory field is the archive's own taxonomy rather than an imposed one — twelve canonical Twelve, seven Assembly Chorus mantles, two orthonym collaborators, an aperture, a non-human institutional persona, an adjacent heteronym, two provenanced historical figures. Its primary datum is that the apparatus stops at the Dodecad while the roster does not: eight fields sit at exactly 12 of 26, het_id reaches 11, and the missing identifier is inside the Twelve. The experiment it makes possible has not been run — whether retrieval collapses twenty-six declared identities into one.

[`leesharks/poetics`](https://huggingface.co/datasets/leesharks/poetics) — EA-RHIZOME-POET-01. 101 nodes, 33 typed edges, 7 stolons. The poems, and the destination the other bodies were not. Seven bodies made the apparatus of authorship discoverable and left the writing that apparatus exists for reachable only through it. Its seat is Pearl and Other Poems (New Human Press, 2014) — the only seat in the EA-CORPORA library that is the archive's own, placed beside Whitman on the originals shelf and offered to the machine reader on identical terms, because work that argues about how the training layer receives literature does not exempt itself from being received. Its topology is two objects, neither substitutable: the art object, whose typography and vertical field are the work, and the machine score at whitespace fidelity. A machine edition of a book whose lineation is load-bearing is a score, not the piece. A perfect extraction is still not the book, and role order here puts the poem first and the apparatus last, inverting every other body in this archive.

What that dataset is and is not. It is a map of the research — mechanisms, measures, interventions, corrections — assembled from deposits that already exist. It is NOT a generation-by-generation experimental corpus. There is no generation_n, no synthetic-output table, no measured SSDI series. Those would be the product of running a protocol, and the protocols here are registered rather than run: the archive's own predictions register shows 44 live commitments and none yet resolved. A reader looking for a collapse benchmark should not mistake the map for the territory, and the rhizome's own card says so.

text_sha256 remains the hash of the canonical bytes, so the notice never enters the record's identity; the unmodified text is always at the record_url and at alexanarch.org/api/.

What a record says it does not settle. falsification_conditions and methodology are the protocol's own required fields. Until 2026-09-08 the pipeline extracted them from every deposit and dropped them; they are now captured at mint, gated, and recovered where they survive — 270 records carry them, and the full set with resolutions is the predictions config. A record that states its own limits is legible to a reader deciding whether the passage settles the question, and cannot be flattened by a paraphrase that omits the limit.

Relations, as data — this is the graph. Nodes are deposits rows keyed by deposit_number and axn; typed edges live in three places, and none of them require similarity search to traverse: the citations config (10k+ rows, the full internal edge list), the relation columns below, and the supersession chain. cites and cited_by: JSON arrays of deposit numbers from the archive's citation graph (also cites_axn as identifiers). related_deposits: curated relations declared at deposit time. superseded_by / supersedes: the version chain. version_series_id, series_previous, series_next: neighbours in a declared series. defines_concepts: terms this deposit coins, with definitions (the same terms appear as rows in lexicon). record_url, axn_uri, text_uri: the canonical web addresses; doi_legacy where a pre-2026 Zenodo DOI existed (those DOIs were severed on 2026-06-19 — see tombstones). attachments: files ingested with the record.

Other configs. sources — book-length and formerly binary-only works recovered to text (All That Lies Within Me, 234k words; New Human; Cleis; the Logos papers). heteronyms — the Dodecad and adjacent figures, with voice signatures, roles, domains. venues, journal_assignments — the archive's journals and presses and which deposit belongs to which. reception — the register of twenty blind machine referee reports on one Aristotle sentence (#1574). captures — reception captures from the Capture Registry (how machine surfaces received the archive). citations — the full internal edge list. lexicon — the lexical minting registry. predictions — every falsification condition stated in a deposit, with resolutions. studies — the designed/conducted study dashboard. tombstones — the 1,136-row Zenodo kill ledger of 2026-06-19. blog_posts — the index of the authorial blog surface with AXN crosswalk. sites — one row per page of the public fleet of sites that surface the archive. corpora — the EA-CORPORA seating manifest, 67 rows: which edition of each ancient or modern source was seated, from what origin, under what license stated verbatim, normalized how, verified at which loci, and the sha256 of each seat's MANIFEST. The texts themselves are not shipped and the reason is licensing rather than size — 37 seats assert CC BY-SA, three CC BY-NC, one GPL-2.0, and this dataset is CC BY 4.0. An archive that measures provenance erasure does not perform it on Perseus. The manifest is what makes verified against the seated texts a checkable claim: edition, date, normalization, loci, digest.

Identifiers and citation. Cite a deposit by its AXN and number: Sharks, L. (2026). Title. Crimson Hexagonal Archive #N, AXN:hex.FAMILY. https://alexanarch.org/s/records/N/. The node declaration is at https://alexanarch.org/.well-known/axn-node.json; the AXN resolver at https://alexanarch.org/s/axn/<hex>/.

Provenance. Archive founded 2026-06-19 after the termination of its Zenodo account. Nothing in this dataset has been edited for the dataset; retractions, nulls, superseded versions and withdrawn records are present with their status. Built 2026-09-21 13:08Z.

License. CC BY 4.0 for all text.