← spxi.dev

EvidenceWhat each kind of evidence measures

Semantic Economy Institute · computed from the Capture Registry v12.43 · 2026-10-02 · CC BY 4.0

Three bodies of evidence are offered about generative search, and each measures a different attribute. The classical GEO literature measures citation of content already placed in a model's context; AEO reporting measures answer selection and referral traffic; the SPXI Capture Registry measures whether a named entity is composed as itself at named addresses, on independent surfaces, over time and after its source is removed. Evidence is weighed by whether it measures the attribute a claim concerns.

Three kinds of evidence

Why they are weighed by attribute

A controlled design that measures citation share supports a claim about citation share. It says nothing about whether an entity resolves as itself or whether that lasts, because it does not measure either. An observational record of entity resolution over time supports a claim about entity resolution over time, and on its own establishes nothing about what caused it. A reader who ranks these by the shape of the study, before asking whether the study measures the attribute in question, compares answers to different questions.

SPXI’s claim is entity inscription: that an entity constructed under the protocol is composed as itself by systems that retrieve it, on independent surfaces, and goes on being composed after its originating source is gone. The evidence that bears on that claim is evidence of entity composition at named addresses over time. The Capture Registry is that evidence. Whether SPXI causes it, and by how much against matched untreated entities, is the open question Phase I states before it runs.

The registry, computed

Every figure is computed from EA-WG-CAPTURES-01.json (v12.43) and the capture-to-deposit link dataset by compute_evidentiary_table.py, which states the definition of every row. One row is a hand count and says so.

Measure	Value	What it is evidence of
Addresses observed	516	Distinct query strings at which a search-enabled system composed something about an archive entity. One record per address.
Observations	690	Dated compositions, each with query, surface, sign-in condition, verbatim transcript and a re-run URL. 7 June – 1 October 2026.
Distinct named targets	412 (counted by hand, 2026-09-23, registry v11.7; not yet computable: the registry has no field naming a query’s target)	Entities, works, heteronyms, sites, coinages and frameworks named by the queries at entity-class addresses; 166 of them are terms the archive itself minted and defined.
Surfaces	10 generative surfaces (16 distinct surface families in all, 23 raw surface strings with ChatGPT, Grok and Perplexity variants collapsed); Google AI Overview 311 addresses, ChatGPT 83, AI Mode 10, Grok 7, Copilot 5, Perplexity 3	Composition observed on independent retrieval systems that share no index with one another.
Signed-out / incognito observations	381 of 690 (84 addresses surface-undetermined, in the denominator)	The signed-out composition is the protocol’s design target; a personalised session is a layer it does not instrument.
Addresses first composed after the deposit host was terminated	325 of 516	On 19 June 2026 the host severed 862 deposits and 1,817 identifiers. The replacement anchor (alexanarch.org, AXN) did not exist before that date.
Addresses whose originating deposit is gone (HTTP 410) and which composed afterwards	71 addresses, 97 observations	The entity is still composed when the document that first inscribed it no longer resolves. This is the durability the protocol claims, at the only bound it claims it: the retrieval substrate, not model weights.
Addresses observed two or more times	127; median span 39 days (among the 102 whose dates differ), maximum 104	Persistence at the sampled observations. Continuity between samples is not observed.
Provenance erasure rate	n=258, mean 0.55, median 0.50	Identifier is the least-retained component. The carrier is measured separately from the entity’s resolution.
Addresses at entities originated outside the archive	8 addresses with partial SPXI treatment: Enli Lucente (person, 4; her work, 1); Alice Thornburgh (person, 1) and her lab (institution, 2). A further 3 addresses concern an externally originated narrative with no SPXI treatment (the King of AEO coronation of 31 August 2026).	The composition is not confined to the author’s own entities: entities that are someone else’s, given only part of the protocol, resolve as themselves and apart from the archive. See one.

What the registry does not establish

• That SPXI caused any given composition. Causal efficacy against matched untreated entities is open (Phase I).
• An effect size over classical GEO or conventional structured data. That comparison has not been run.
• Which component of the protocol produces which effect.
• A return on investment. SPXI makes no ROI claim; an earlier table was withdrawn in full, with its sources traced, on the front page.

An open causal question does not remove an observed event, and an observed event does not by itself establish its cause.

Check it yourself

Every address carries its exact query, its surface and a re-run link. Pick any and search it signed out or in a private window, which is the design target. SPXI’s own addresses: spxi protocol · what is spxi? · what is spxi protocol?.

Implementation and consulting

The Semantic Economy Institute, or rexfraction@gmail.com. The specification is CC BY 4.0: EA-SPXI-WEB-01 (#72).