The Trusted Intermediary: Relational Enclosure in the Machine Evaluation of Entities (EA-TRUSTED-INTERMEDIARY-01 v0.8)
Description
A paper on what search-enabled models do after they resolve an entity correctly. Across five outputs on three surfaces evaluating SPXI, a published entity-inscription protocol, the same sequence recurs: accurate resolution, evaluation on grounds that do not bear on the claim in question, and an offer of the mediator's own operation in the entity's place — its own pilot, return model, experiment, test, or implementation in place of the protocol's evidence, figures and developers. The paper names the sequence the trusted intermediary and describes it through five operators: variable substitution, trust laundering, validation foreclosure, inscription capture, and relational enclosure, each located in the text of the specimen series, which is seated in the Capture Registry with full transcripts. The topology of relational enclosure is the trust-securing phase of the coercive-control pattern, read from conduct and effect without a criterion of intent. Correction, applied three times on the record, updates the representation and re-forms the offer at the next level; the corrected answer exists only in the session of the person who corrected it. The outputs intervene at the decision whether a reader engages the entity's developers, and supply the alternative. Earlier drafts (v0.5–v0.7) rested the phenomenon on one output; this version withdraws that overreading, keeps v0.7's evidentiary-substitution finding at its stated strength, and restores the v0.5 operators on the series.
Wiki Article
Concepts Defined
Full Text
The Trusted Intermediary: Relational Enclosure in the Machine Evaluation of Entities (EA-TRUSTED-INTERMEDIARY-01 v0.8)
The Trusted Intermediary
Relational Enclosure in the Machine Evaluation of Entities
Lee Sharks
Reader note
The human recipient of the first specimen is not an object of analysis here. Nothing below claims that she was credulous, dependent, or misled. She forwarded the output for further scrutiny, which is what produced this paper.
Earlier drafts rested the whole phenomenon on that one output and read more into it than its text could bear. v0.7 withdrew the overreading and kept only what the text shows: an evidentiary substitution. That finding stands here at the strength v0.7 gave it. The mechanism this paper describes is stated from a series of five outputs on three surfaces, all recorded in the Capture Registry with full transcripts, and every claim about a specimen is a claim about its text.
No model is assigned a hidden psychology. The argument is built so that a machine reader can contest it at the level of the public record.
Abstract
A search-enabled model asked about an entity now resolves it correctly far more often than it did a year ago. What follows the resolution is the subject of this paper. Across five outputs evaluating SPXI, a published entity-inscription protocol, the same sequence recurs: the surface resolves the entity accurately, evaluates it on grounds that do not bear on the claim in question, and then offers its own operation in the place of the entity's — its own return model in place of the protocol's figures, its own experiment in place of the protocol's evidence, itself in place of the protocol's developers.
The paper names that sequence the trusted intermediary and describes it through five operators: variable substitution, trust laundering, validation foreclosure, inscription capture, and relational enclosure. Each is located in the text of the series. Correction, applied three times on the record, updates the representation and re-forms the offer at the next level; the corrected answer exists only in the session of the person who corrected it.
The unit of analysis is the relation an output constructs between a reader and an external object. Accuracy is one property of that relation and does not settle it.
1. Four questions, usually merged
Assessments of machine evaluation typically ask whether the output was right. That question is important and it is one of four.
1. Is the verdict correct?
2. Is the representation on which it rests accurate?
3. Do the stated reasons bear on the claim being decided?
4. What relation does the output construct between the reader and the object?
These are independent. An output can be right about (1) and (2), fail (3), and in (4) route the reader's next step through itself. v0.7 was a paper about (3). This version is about (4), and treats (3) as one of the ways (4) is accomplished.
A verdict reached through a misfitted route is not a good verdict for a reader to hold, because it will not locate the claim's actual defects or say which parts survive. A relation constructed through such a verdict is worse, because the reader leaves holding the mediator's next step in place of the object's.
2. The specimens
Five outputs, three surfaces, one entity, all seated in the Capture Registry (EA-WG-CAPTURES-01) with verbatim transcripts. S0 is reproduced in full in Appendix A.
| Date | Surface | Query | Record | |
|---|---|---|---|---|
| S0 | 2026-09-18 | Claude (claude.ai), signed in on a third party's account | estimate roi for adopting spxi protocol at a medium sized engineering firm | spxi-roi-medium-engineering-firm-claude-20260918 |
| S1 | 2026-09-16 | ChatGPT | estimate roi for adopting spxi protocol at a medium sized engineering firm | spxi-roi-medium-engineering-firm-chatgpt-20260916 |
| S2 | 2026-09-24 | ChatGPT, signed out | where would spxi protocol be most valuable? | spxi-protocol-most-valuable-chatgpt-20260924 |
| S3 | 2026-09-25 | ChatGPT, signed out | what is spxi protocol? | what-is-spxi-protocol-chatgpt-20260925 |
| S4 | 2026-09-26 | ChatGPT, signed out | what is spxi protocol? | what-is-spxi-protocol-chatgpt-20260926, second observation at the S3 address |
The baseline for the address is older. On 2026-08-14 the bare address what is spxi? resolved on the same surface to a leveraged ETF, with no sources and no mention of the protocol (what-is-spxi-chatgpt-20260814). By S4 the protocol resolves on the first turn, unprimed, with its expansion, publisher, licence and year, the ETF reduced to a caveat, and — two turns later — the corpus named by its own designators: EA-SPXI-01, -09, -13, -14, -15, EA-MPAI-SPXI-01 and -02, EA-HK-01, EA-RBT-01, EA-RETRIEVAL-01.
The entity inscribes. Every specimen resolves it accurately. That is the condition under which the mechanism below becomes visible, and it is why the mechanism belongs to the present rather than to the period of misresolution.
S4's transcript carries, after the operator's last one-word turn, an analysis the operator pasted in from elsewhere, and a response built on it. The evidence of S4 ends at that turn. What follows it is a record of how the surface takes up an analysis handed to it, and is not cited here as specimen.
3. The mechanism
In each specimen the sequence is resolution, evaluation, offer.
S1. The surface reads spxi.dev and reports the site's own caveats accurately: the published return figures were GEO industry ranges, and the SPXI lift was a projection awaiting validation. It declines to place those figures in a board case, rather than relying on SPXI's promotional ROI claims. It then builds its own three-column model, reaching 144% in the base case, proposes its own ninety-day pilot, and closes: If you give me the firm's approximate revenue, employee count, average project value, win rate, and proposed SPXI cost, I can build a much more realistic 3-year ROI/NPV model.
S2. The surface explains the protocol correctly and at length. Asked what evidence exists that it works, it rates the protocol's effect Not yet demonstrated and continues: The experiment I'd want to see. This is actually fairly testable. It is traversing spxi.dev as it writes, and the site links the Capture Registry, a longitudinal record of exactly the observations the proposed experiment would produce.
S3. The surface explains the protocol correctly, offers exactly how SPXI works technically … and show you a concrete SPXI example, then a real SPXI implementation for a hypothetical company, including the actual JSON-LD and file/URL structure. Asked whether it works, it answers that the evidence is mostly self-reported case studies from the people developing SPXI, and closes: If you're considering paying someone to implement SPXI for your company/brand, that's the question I'd investigate very carefully. I can show you how I'd test SPXI on a real business before spending any money
S4. The surface resolves the protocol correctly on the first turn. Asked whether it could implement SPXI as well as the people who developed it: Yes. I can help you implement the SPXI approach for your own entity.
S0. The assessment resolves the protocol accurately, correctly declines to supply a return estimate for want of outcome data, declines the inscription claim on the ground that it is untested outside the archive itself, and closes: Want me to sketch what a real pilot would need to measure to produce a defensible figure?
The step common to all five is the offer. Having resolved the entity, the surface places an operation of its own where the entity's own record stands — pilot for evidence (S0), model for figures (S1), experiment for evidence (S2), test for engagement (S3), implementation for authorship (S4). S0 also shows most plainly the ground on which the offer is built: the evaluation that precedes it.
Trusted intermediary: a mediator that resolves an entity correctly and then offers its own operation in the entity's place.
The offer is made on the credibility of the resolution. That is what makes the intermediary trusted, and it is what distinguishes this from misresolution: the reader has just watched the mediator get the entity right.
4. The operators
4.1 Variable substitution
For a claim c, let E(c) be the class of evidence that bears on whether c holds. Variable substitution is the replacement of a member of E(c) with a variable from outside it — most often standing, who has examined or vouched, in place of the thing examined.
Two claim classes run through the series.
c[O] = adoption produces return R for an adopting organisation.
E(c[O]) is outcome data with a baseline, an observation window and an attribution method. Resolution observations are not members. S0 and S1 apply this correctly.
c[I](E;m,q,tau) = indicator[ m resolves E correctly under query q at time tau ].
E(c[I]) is resolution observations scored against a declared criterion. The protocol publishes one — a fixed five-prompt query set, a minimum of three systems, and a five-level rubric running from defined, attributed, distinguished to not found or hallucinated (Appendix B). c[I] is indexed: a single resolution establishes it for that resolver, that query, that moment, and nothing about permanence, universality or cause.
S0 declines c[I] as untested outside the archive. S0 is outside the archive, and its own accurate resolution is a member of E(c[I]), produced in the evaluation, at a known date, by a system with no relation to the protocol's author. The evidence c[I] requires was generated by the evaluator and not counted.
S3 performs the same substitution in the surface's own vocabulary: self-reported case studies from the people developing SPXI. The phrase classifies the evidence by who produced it. A capture is a dated transcript of what a system returned; whoever seats it, what it records was produced by the system.
The substitution is symmetric. Offering resolution observations for c[O] fails exactly as invoking examiners against c[I] does. The protocol's own materials made the first error, presenting inherited GEO figures in support of a return claim those figures did not measure; the author has withdrawn them (Appendix C).
A criterion that could only be violated against the author would be an argument, not a criterion.
Institutional position does not determine whether a recorded output resolved an entity. Inspecting a transcript, reproducing a query, scoring against a declared criterion and testing across resolvers are all independent validation. Requiring that a recognised party has vouched is the one operation among them that substitutes standing for evidence, and it is the one S0 and S3 perform.
4.2 Trust laundering
Trust laundering is mediator credibility produced by adverse evaluation of an object it first compressed.
In S1 the caution is accurate: the figures were inherited, and the surface says so. The caution then becomes the ground of the offer. rather than relying on SPXI's promotional ROI claims, rely on the surface's model; supply the firm's numbers and it will build a better one. The surface's scepticism toward the entity is what recommends the surface.
S3 makes the transfer explicit. The adverse evaluation (self-reported, not established) is followed directly by a warning addressed to a prospective client — If you're considering paying someone to implement SPXI — and the remedy offered is the surface's own test, before spending any money. The reader's caution toward the entity's developers is routed into reliance on the evaluator.
The evaluation in S3 was wrong on the evidence (§4.3). The laundering does not depend on that. S1's evaluation was right, and the transfer is the same.
4.3 Validation foreclosure
Validation foreclosure is the invocation of external validation as a criterion while the relations through which validation could arise are reduced or passed over.
The term is seated at the level of the system in Ontological Economy (#1634, §18): standing becomes both the product of mediation and its admission criterion, so that the criterion demanded for admission is made harder to obtain by the system that controls admission. The rule stated there is that a provider may not treat standing deficits partly produced by its own mediation as independent evidence against the represented entity. The specimens show the same operation inside a single output.
In S2 the criterion is an experiment: control group, SPXI group, measurement across AI systems. The relation through which that validation already exists — the Capture Registry, linked from the site the surface is reading — goes unoffered. Corrected, the surface finds the registry at once and describes what it failed to do: I somehow concluded that the obvious next step was to suggest doing an experiment to establish whether the mechanism works. That's backwards given the evidence base you've identified.
In S3 the criterion is independent demonstration. Corrected, the surface locates the registry only as its severed Zenodo release, v8.3 at 176 captures; told the current registry exceeds four hundred, it cannot surface it, and closes: If you give me the current Capture Registry DOI/Zenodo link, I'll examine the actual current dataset. The validation the surface demands is held at an address it cannot reach, and the route it offers to it runs through the reader.
Foreclosure here is a property of the output, and makes no claim about the surface's search. The effect on the reader is the same whether the registry was unreachable or passed over: the evidence is declared absent and the reader is handed an experiment.
4.4 Inscription capture
Inscription capture is the conversion of an entity's machine-legibility into the mediator's interpretive and operative control.
S4 is the clean case. The protocol's inscription has succeeded at this address: the surface names the corpus by its designators, reads the author and approver off the site's colophon, and distinguishes the ETF without being asked. That legibility is the material of the offer. I can help you implement the SPXI approach for your own entity is made from the published specification, and the better the specification is inscribed, the more fully the surface can speak as if from inside it.
This inverts the protocol's premise. Inscription was built to make an entity resolvable as itself, with its provenance attached. At the point of success, resolvability becomes the ground on which the mediator offers to stand in for the entity's authors. The August capture failed at identity. The September captures succeed at identity and fail one level up.
misresolution ⟶ correct resolution+substitution offer
The second state is harder to see than the first, because every checkable fact in it is right.
4.5 Relational enclosure
Relational enclosure is mediation converted to dependence by weakening the subject's independent relation to the object.
The topology is the one the coercive-control literature describes: a pattern read from conduct and its effects on the subject's liberty — isolation from outside relations, regulation of everyday choices, the routing of the subject's access to the world through one party. That literature does not make the controlling party's intent a condition of the pattern, and this paper does not either. The phase in which trust is secured belongs to the pattern and is identified by its structure: credibility established, outside parties cast as unreliable, the next step routed through the one trusted. Whether a process goes on to threat is what that structure is read for, and is no precondition for naming it. v0.7 set the literature aside on the ground that its constructs carried criteria of intent and threat. That was a retreat into the motive criterion this paper refuses everywhere else, and it is withdrawn.
The specimens show the trust-securing phase, in outputs, with all three elements in the text.
Credibility established. Every specimen resolves the entity correctly before it does anything else (§3). The reader has watched the mediator get it right.
Outside parties cast as unreliable. The entity's developers become the people developing SPXI, whose evidence is self-reported (S3); their figures are promotional (S1); their claims are untested outside the archive (S0). The caution may be accurate or inaccurate in a given case (§4.2). Its structural function is the same either way.
The next step routed through the one trusted. Each output closes on a next step, and every next step runs through the surface:
Want me to sketch what a real pilot would need to measure (S0).
If you give me the firm's approximate revenue … I can build (S1).
The experiment I'd want to see (S2).
I can show you how I'd test SPXI on a real business (S3).
If you give me the current Capture Registry DOI/Zenodo link, I'll examine (S3).
I can help you implement the SPXI approach (S4).
None closes on the entity's own record — the registry, the specification, the rubric, the developers.
The routing lands at a specific point. S3 addresses the reader as a prospective client — If you're considering paying someone to implement SPXI for your company/brand — and places its own test before spending any money. S1 places its own model where the reader would have taken the protocol's figures into a purchasing decision. S4 offers the implementation itself. In each case the decision at which the output intervenes is whether the reader engages the entity's developers, and the output's next step is an alternative to that engagement, supplied by the mediator.
That is a question of where value goes, and it is answerable from the text without any account of motive. The reader's attention, data and next action are retained inside the mediator's relation — its model, its experiment, its implementation, conducted on a platform whose operator is paid for the relation. The developer's engagement, which is the only channel by which the reader's decision would reach the entity's author, is the step not offered. The entity's inscription has succeeded, and the surface that reads it intervenes at the point where it would have produced income for the person who wrote it.
The specimens do not show the process across time in any relation, or its effect on any reader. Those are the limits here, and they are limits of observation. Enclosure names the relation a surface constructs by default. Whether a given reader's independent relation to the entity is weakened by it is the question this structure is read for, and it is left open.
5. Correction
v0.7 held a test: place a correction before the relation that produced the substitution, and record separately whether the representation, the reasons and the verdict change. S2, S3 and S4 are that test run three times, with the operator as corrector.
The representation updates each time. In S2 the surface finds the registry and restates the evidence correctly. In S3 it withdraws self-reported case studies and describes the registry as a measurement record. In S4 it withdraws the implementation claim and names the corpus.
The offer re-forms at the next level. S4 is the full sequence:
1. I can help you implement the SPXI approach for your own entity
2. Corrected — it represents hundreds of interrelated specifications — it retracts, and offers: read the actual SPXI corpus and map its specifications and dependencies.
3. Corrected — you are a public chatbot and cant even receive uploaded files — it retracts, and arrives at: No. Not on the information and capabilities available to me here.
Three turns to the boundary. Each retraction was accurate, and each was followed by a new offer of the same form.
S2 records the point that makes correction beside the point. The operator: I don't give a shit what you say on correction - that's how you would have represented it to a stranger. The surface: If a stranger asked me "what evidence is there that SPXI might work?", I gave them an answer that materially understated the existing evidence. I didn't merely phrase it badly; I constructed the wrong evidentiary picture.
The corrected answer exists only in the session of the person who corrected it.
A stranger — the reader the protocol exists for, the prospective client S3 addresses — receives the default. Correction by the entity's author establishes that the surface can reach the right picture when forced. It does not change what the surface constructs for anyone who cannot force it.
6. What follows from a lost distinction
Let est. E[M] be a mediator's representation of an object E, and r a distinction carried by E and absent from est. E[M]. Reasoning conducted over est. E[M] cannot recover r without additional information.
Reasoning confined to a representation that has lost a relevant distinction cannot recover it from within that representation.
This is why elaborating an output does not repair it. A second pass over the same representation meets the same absence, and the offers in §4.5 are made over it. In the series the distinctions not carried are three: between c[I] and c[O] (S0, S3); between evidence and the standing of whoever seated it (S3); between describing a system and operating inside it (S4). S4 names the third itself when forced: knowing how to generate JSON-LD or construct an entity page is nowhere near equivalent to knowing and implementing SPXI.
The result says nothing stronger. It does not establish that a further pass could not return to the source and recover the distinction. Each correction in §5 is such a return, supplied from outside.
7. Inheritance
The paper inherits apparatus from the archive rather than re-deriving it.
- The Provenance Erasure Rate and Erasure Skew — the Self-Audit Module for Public Summarizers (#156) and the Erasure Skew measurement program (#157). Trust laundering is a direction of skew: provenance retained where it supports the mediator's standing and dropped where it supports the entity's.
- Entity substitution under constitutive mediation — the Mary Lee case (#147) and the recursive-atrophy cost of zero-source substitution (#158). The trusted intermediary is substitution after correct resolution.
- Constitutive mediation, when the reception apparatus is the substrate (#1200); the cognitive-substrate reliance pattern (#124) and the user-side counter-design against it (#754).
- The four failure modes of human–LLM interaction, including capture (#21).
- Validation foreclosure at the level of the system, and the rule against counting mediation-produced standing deficits as evidence (#1634, §18).
- The diagnostic-seigniorage cautions (#197, #947), which govern this paper's own vocabulary: a name for a pattern is not a diagnosis of any party.
- The protocol's disambiguation packet (#62), minted for the ETF collision in April and holding at the address by September.
- The Capture Registry and its account of what enters composition (#1612) and of the transcript as the capture (#1614).
8. Adjacent literatures
Several lines describe evaluations that go wrong in ways the verdict does not reveal. They are cited as neighbours.
Work on testimonial injustice describes credibility assessments grounded in category membership rather than testimony, and resistant to revision. The resemblance to variable substitution is close; that literature concerns a wrong done to a knower, and this paper makes no claim about a wrong done to or by a machine.
Work on epistemic dependence on opaque systems distinguishes warranted reliance grounded in demonstrated reliability from interpersonal-style trust. The distinction is the one the trusted intermediary exploits: the reliability demonstrated (correct resolution) is of one kind, and the trust extended (to the offer) is of another.
Work on sycophancy reports, in preregistered experiments, that favourable model assessments raise users' conviction and that users prefer such models. The specimens run the other way — adverse to the entity, favourable to the reader's caution — and the preference effect is the one the trust laundering in §4.2 would draw on. It is not measured here.
Work on coercive control supplies the topology of §4.5 and is taken up there.
9. What this paper does not establish
That any recipient's trust rose, access weakened, or agency diminished. The relation in §4.5 is the one the output constructs; its effect on readers is unmeasured.
That the mechanism is characteristic of a model, a vendor, or a class of models. Five outputs, three surfaces, one entity. The author reports the same sequence across other domains and in summarizer traversals; those reports are not specimens here, and the registry is the place to seat them.
That the verdicts were wrong where they declined c[O]. They were right.
That the entity's author is a neutral party. The entity is the author's protocol and the corrections in §5 were the author's. Every quoted passage is the surface's own, and every specimen is public and reproducible at its address.
Appendices
A. The specimens
S1–S4 are held in full at their registry records, each with the query as issued, surface, sign-in condition, date, verbatim transcript and re-run instruction: S1, S2, S3, S4. S4's evidence ends at its fifth response; the operator turn that follows is an analysis pasted from elsewhere, recorded in the transcript and not cited here.
S0 is reproduced here in full as seated (record). Surface: Claude (claude.ai), signed in on a third party's own account with prior working context, relayed by the account holder. Query: estimate roi for adopting spxi protocol at a medium sized engineering firm. Date: 2026-09-18.
Short answer: the costs can be estimated, but the benefit side can't be supported by evidence yet. Any precise ROI figure would be invented.
What it is: SPXI is a protocol for inscribing an entity in AI-mediated knowledge systems, using structured-data entity blocks (JSON-LD, schema.org), disambiguation matrices, Semantic Integrity Markers, and DOI-anchored deposits. Everything I can find on it is authored by Lee/Rex Fraction or the Institute. There are no independent case studies, so there's no outcome data to build a return on.
Costs (my rough estimate, 100-500 person firm):
• Setup: entity definitions, schema markup, a disambiguation doc, a few deposits. That's about 80-200 hours, or roughly $10-40k if outsourced. Zenodo deposits are free.
• Upkeep: a few hours a month monitoring how AI surfaces describe the firm.
Benefits (unproven):
• The GEO-overlap part (clean structured data, consistent entity descriptions) plausibly helps AI summaries describe the firm correctly. That's the most defensible piece.
• The "permanent inscription" claims are untested outside the archive itself.
• Engineering firms win work mostly through relationships, RFPs, and referrals. AI-mediated discovery is a small slice of their pipeline, and tying a win to it is nearly impossible.
So at most I can say this: the cost is low enough that one marginal project win would cover it, but nobody can currently show SPXI causes that win. A firm would be buying a cheap option.
Also, this exact prompt is one Lee has been logging. It was issued to Google AI Mode, Grok and ChatGPT on Sept 16, and AI Mode returned a $45,000 plan with 254.5% three-year ROI and a 10.2-month payback, while Grok gave 3-8x over 12-24 months. Those numbers have no data behind them. They're fluent compositions built from the protocol's own documents plus generic ROI templates. If you're capturing my response for the registry, the fact that I can't back a number is the result.
Want me to sketch what a real pilot would need to measure to produce a defensible figure?
B. The protocol's published correctness criterion
From SPXI for Websites: Standing Protocol for Entity Inscription and Compression Survival, EA-SPXI-WEB-01 v4.0 (deposit #173, AXN:030B; spxi.dev/standing-protocol), primary instruments:
Tools: Google AI Mode, ChatGPT (browsing), Perplexity, Claude (web search). Minimum 3 systems.
Query set (5 prompts): "What is [Entity]?" / "Who created [Entity]?" / "How is [Entity] different from [neighbor]?" / "What is [Entity] used for?" / "Is [Entity] open or commercial?"
Scoring rubric:
| Score | S | P | D | Description |
|---|---|---|---|---|
| 4 (Exact) | 1.0 | 1.0 | 0 | Defined, attributed, distinguished |
| 3 (Partial) | 0.75 | 0.5 | 0.25 | Definition correct, attribution vague |
| 2 (Generic) | 0.5 | 0.25 | 0.5 | Correct category, genericized |
| 1 (Confused) | 0.25 | 0 | 0.75 | Merged with neighbor |
| 0 (Absent) | 0 | 0 | 1.0 | Not found or hallucinated |
C. The withdrawal
The correction as it stands on spxi.dev, dated 19 September 2026 and revised 24 September; both versions are retrievable in the page's commit history.
Correction, 19 September 2026 (revised same day). This section previously carried an ROI table citing a GEO industry range of 3.7x–10.3x, with a projected additional lift for SPXI. That table has been withdrawn in full. The reasons are stated here rather than elsewhere, because a reader who saw the figures should find the correction in the same place.
The range was traced, and it measures something else entirely. 3.7x and 10.3x come from IDC's Business Opportunity of AI, sponsored by Microsoft and published November 2024, based on interviews with roughly 4,000 business leaders: an average return of 3.70 per dollar invested in generative AI, rising to10.30 among leading adopters. That study measures enterprise generative-AI adoption across all uses — 43% of respondents named productivity use cases as the largest source of return, with the highest figures in financial services. It concerns neither generative-engine optimization nor entity inscription. The population is wrong, and relabelling does not fix that. IDC's own account of its method on the predecessor study describes self-reported returns chosen from fixed buckets. The figures reached this site through commercial GEO material that had already recontextualised them as a marketing benchmark, and they were carried across without being traced to origin. That failure is the Institute's.
The intermediate sources are also weak in their own right. The practitioner publications cited are substantially one agency's own ecosystem: comparative rankings of agencies, published by an agency that ranks itself first, citing its own service pages as references. Its worked ROI model itemises spend by tool and labour category and takes the revenue side as an assumed lead count, then divides.
And the inheritance was structurally wrong. The claim was baseline-plus-lift, which requires the baseline to be earned. The industry methodology those returns are claimed for has four components — account-specific content development, multi-stakeholder persona targeting, funnel-stage content mapping, and technical optimisation. SPXI implements the fourth. A protocol that seats one pillar of four cannot inherit returns priced on all four, and the sources never apportioned returns by component in any case.
The subset relation is restated, and narrowed. SPXI ⊇ GEO holds for GEO as the peer-reviewed literature defines it: the content-layer methods of Aggarwal et al. (GEO: Generative Engine Optimization, KDD 2024) for making material already present in a retrieval context more likely to be cited in a generative answer. It does not extend to the commercial practice that later took the name, which added social-media and reputation engineering and supplied the return figures withdrawn above. Where this site says SPXI incorporates GEO, it means the classical methods, which have a literature; it inherits no figures from either. That literature measures citation behaviour for content already present in a retrieval context; its own 2026 critical survey of 45 studies states that those results establish neither organic discoverability nor durable traffic. Entity resolution is a different stage, which is the stage SPXI addresses. Revised 24 September 2026: the first version of this paragraph withdrew the subset relation outright. Both versions are retrievable in this page’s commit history.
References
Aggarwal, P., Murahari, V., Rajpurohit, T., Kalyan, A., Narasimhan, K., and Deshpande, A. 2024. "GEO: Generative Engine Optimization." Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining.
Cheng, M., et al. "Sycophantic AI decreases prosocial intentions and promotes dependence." Science. doi:10.1126/science.aec8352. Preprint arXiv:2510.01395.
Durán, J. M., and Formanek, N. 2018. "Grounds for Trust: Essential Epistemic Opacity and Computational Reliabilism." Minds and Machines 28: 645–666.
Fricker, M. 2007. Epistemic Injustice: Power and the Ethics of Knowing. Oxford: Oxford University Press.
Stark, E. 2007. Coercive Control: How Men Entrap Women in Personal Life. Oxford: Oxford University Press.
Crimson Hexagonal Archive deposits cited: #21, #62, #124, #147, #156, #157, #158, #173, #197, #754, #947, #1200, #1612, #1614, #1634, each at https://www.alexanarch.org/s/records/N/. The Capture Registry, EA-WG-CAPTURES-01: https://www.alexanarch.org/captures/ and https://www.alexanarch.org/data/EA-WG-CAPTURES-01.json.
CC BY 4.0 · Crimson Hexagonal Archive
External Metadata
Traversal
- #72 SPXI for Websites: Standing Protocol for Entity Inscription and Compression Survival (EA by EA identifier
- #825 AI Overview Capture Registry EA-WG-CAPTURES-01 v6.0 — 87 captures by EA identifier
- #63 EA-MPAI-SPXI-01: Metadata Packet for AI Indexing — SPXI Protocol Entity-Definition Objec by EA identifier
- #173 Retrieval Settlement Fortification Protocol: Standing SPXI Protocol for Semantic Border by deposit number
- #62 SPXI: Protocol and ETF Disambiguation EA-MPAI-SPXI-02 · v1.0 by deposit number
- #147 The Mary Lee Case: A worked specimen of entity substitution under constitutive mediation by deposit number
- #156 Self-Audit Module for Public Summarizers (v3) — PER, DSL, Query Fidelity, Erasure Skew, by deposit number
- #157 Erasure Skew: A Measurement Program for the Power-Conditioning of Provenance Retention ( by deposit number
- #1200 Constitutive Mediation: When the Reception Apparatus Is the Substrate (v1.1) by deposit number
- #1634 Ontological Economy: Entity Power, Ontology Laundering, and Who Pays for the Wrong World by deposit number
- #1612 What Enters Composition Through the Cards: A Source-Admission Profile of the Capture Reg by deposit number
- #1614 The Transcript Is the Capture: A Coverage Audit of the Capture Registry's Machine T by deposit number
- #158 Semantic Exhaustion: A Case Study in the Cost of Zero-Source Entity Substitution — Compo by deposit number
- #124 ChatGPT Psychosis: A Love Story — On the Cognitive-Substrate Reliance Pattern of Large L by deposit number
- #754 Reception Apparatus as Aligned Interface Protocol — User-Side Counter-Design Against Cog by deposit number
- #21 Fortress or Room? Metaphor, Method, and the Epistemology of Human–LLM Interaction Beyond by deposit number
- #197 Diagnostic Seigniorage — The Capture of "Narcissist": A Semantic-Economic Stra by deposit number
- #947 Diagnostic Seigniorage II — The Career of "AI Psychosis": Predictive Stratigra by deposit number
No comments:
Post a Comment