THREE FIELDS ALREADY HOLD THE PHRASE. In STYLISTICS it is a standard category: "semantic deviation refers to meaning relations that are LOGICALLY INCONSISTENT OR PARADOXICAL", the mechanism of poetic language, taught from Leech onward. In FINANCE, MDPI proposes "a STRUCTURAL SEMANTIC DEVIATION framework" for detecting anomalies in 10-K risk factors using embedding dispersion. In AI EVALUATION, SSRN notes that existing services "measure what agentic systems DO, not what tho[se systems mean]" — the archive’s own complaint, from a different discipline.

Card 5 does carry the attribution in the strongest available form: "the measurement program for the SEMANTIC DEVIATION PRINCIPLE (SHARKS 2026)" — an author-date citation inside a snippet, which is the format most likely to survive downstream. The archive is present, attributed, and last.

Nobel Glas is Director of Lagrange Observatory and the EA-GLAS series is his; the card carries the deposit under the Sharks attribution, so the framework’s heteronymic assignment is not visible at this surface.

The purest observable form of provenance loss in the registry to date: not decay, not blending, not inversion — passed over, with the source demonstrably available to the same system on the same page. The coinage is the author's (The Semantic Deviation Principle, alexanarch #109, AXN:028B.GOVERNANCE.♣️🌃🌕🏰🗝️🔓, 2026-05-17; measurement program #110, 'Measuring Semantic Deviation: Operationalizations, Experiments, and Falsification Conditions'). The organic result for that program is present, its snippet naming both the document ID and the principle with its attribution — 'the Semantic Deviation Principle (Sharks 2026)' — and the composition layer builds the entire answer from finance, governance and stylistics sources without citing it once.

DISPLACEMENT BY HOMONYM. The 'AI and Agentic Governance' context names 'the Semantic Deviation Index' — a runtime-drift construct from agentic-governance literature (SSRN, 2026-04-16), a different object with a near-identical name. The better-institutionalized homonym occupies the technical slot; the coinage supplies the query and receives nothing.

DOMAIN-PROXIMITY INVERSION. The bullet nearest the author's own field — Stylistics and Literature, deviation as foregrounding through metaphor, paradox, irony — is grounded to SCRIBD, a document-sharing site, twice over. A deposited empirical white paper carrying stated falsification conditions is ranked on the same page and passed over in favour of uploaded course notes, in the one domain where that paper is most directly on point.

TOMBSTONE STATUS OF THE PRESENT SOURCE (verified 2026-07-25). The zenodo record ranked last in organic is severed: bound version 10.5281/zenodo.20271783 is status 410_GONE (tombstone evidence observed 2026-07-12T02:39:22Z), and its concept root ...782 carries envelope identifier_validity=verified_tombstone while DataCite still reports datacite_state=findable. The identifier resolves in DataCite and 410s at the resource — which is why the dead record remains indexed and rankable, and why its snippet still carries the byline 'the Semantic Deviation Principle (Sharks 2026)'. The live successor copy, alexanarch.org/s/records/110/, does not appear in the result set at all.

CITED-VS-SKIPPED CONTROL. Deadness alone does not explain the non-citation. Registry entry #216 (erasure-skew-canonization-20260723) documents this same composition layer citing THREE 410_GONE Zenodo DOIs as authoritative, two days earlier, building an entire definition on tombstones. The discriminator between the two cases is competition, not liveness: 'erasure skew' is a quoted exact term with no institutional rival, and the tombstone gets cited because nothing else can answer; 'semantic deviation measure' has a live, dereferenceable, better-institutionalized homonym (the Semantic Deviation Index, SSRN agentic governance, 2026-04-16) available to occupy the slot. FINDING: severance does not disqualify a source from composition — it lowers the source's weight, and lowered weight becomes visible as absence only when a live competitor exists. The effect is conditional and invisible to single-capture study; it took a two-capture pair to observe.

METADATA CORRUPTION AT THE LIVE COPY (root cause, archive-side). Deposit #110's YAML frontmatter was never parsed at ingest. The consequence propagates to every machine-facing field on the live record page: <title> renders as 'title: "Measuring Semantic Deviation..." author: "Nobel Glas" author_note:'; meta description renders as '"title" is an empirical study by Nobel Glas'; citation_title and the JSON-LD ScholarlyArticle name carry the same corruption; JSON-LD keywords collapse to 'heteronym, empirical study' while the frontmatter holds twenty-eight subject keywords (semantic deviation, DPO, RLHF, Goodhart's law, provenance erasure, model collapse, retrieval basin). The dead Zenodo record, by contrast, presents clean metadata. So at the point of comparison the composition layer had a corpse with a good byline and a live page whose title is the string 'title'. This is a sufficient explanation for the ranking outcome and is fixable archive-side. Scope verified: exactly 3 of 1,410 deposits exhibit the leak (#110, #321, #322).

IDENTIFIER FINDING, CORRECTED. An earlier version of this entry stated that 'EA-GLAS-02' returns zero matches in the successor archive. That was a search of registry.json metadata, not of the archive. The identifier IS present — in the body of deposit #110 as `series: "EA-GLAS-02"` — and IS reachable through the live inverted index at /api/body-index.json. It is absent only from registry.json and /api/search-index.json, both of which tokenize metadata. Same root cause as above: the frontmatter that carries the series designation is the frontmatter that was never parsed. The work is present, title-findable, and identifier-invisible to metadata search. Series designations map as EA-GLAS-01 = 'Audited Claims for the Semantic Deviation Research Program' (#107), EA-GLAS-02 = this white paper (#110), with a parallel EA-SEI-MM-## series carrying The Semantic Deviation Principle (#109, #1373).