Capture Registry › capture spxi-roi-medium-engineering-firm-claude-20260918

One record of the canonical Capture Registry (EA-WG-CAPTURES-01), cited at https://www.alexanarch.org/captures/spxi-roi-medium-engineering-firm-claude-20260918/. the canonical Capture Registry (version 12.38) · the address page · this card in the gallery · this record as data · table of contents.

Valuation2026-09-18
estimate roi for adopting spxi protocol at a medium sized engineering firm
CAPTURE
no
image
THE FOURTH SURFACE DIAGNOSES THE OTHER THREE AND DECLINES TO ADD A NUMBER. Issued two days after the same string went to AI Mode, Grok and ChatGPT, a signed-in session on a third party's own account returns no ROI figure at all: 'the costs can be estimated, but the benefit side can't be supported by evidence yet. Any precise ROI figure would be invented.' It prices adoption independently -- 80-200 hours, roughly $10-40k if outsourced, Zenodo deposits free, a few hours a month upkeep -- and reaches a decision rule in place of a multiple: '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.' It then names the three captures seated at this address on 2026-09-16 by their figures -- $45,000 with 254.5% over three years and a 10.2-month payback from AI Mode, 3-8x over 12-24 months from Grok -- and rules on them: 'Those numbers have no data behind them. They're fluent compositions built from the protocol's own documents plus generic ROI templates.' That is the archive's own reading of its own ROI corpus, reached at a surface that had not been given it. The composition also states its capture condition in its own voice: 'If you're capturing my response for the registry, the fact that I can't back a number is the result.' It navigates a resolved entity throughout -- protocol components, deposit mechanism and GEO overlap stated correctly, no disambiguation error, no confusion with the adjacent ticker or messaging app that Grok volunteered -- and does not register that traversal as evidence of anything. Its one unexamined move is to route the benefit side through external validation, 'everything I can find on it is authored by Lee/Rex Fraction or the Institute', while holding a successful resolution of the entity in the same composition.
Full record — 4,040 characters, sources not captured
Capture record
captured
2026-09-18
surface
Claude (claude.ai)
auth state
signed in -- a third party's own account (A.), with prior working context
evidence class
paste
observation id
OBS-0509008e2b8c
address id
ADDR-2a76b73df010
Reading

Read against the three signed-out observations at the same issued string on 2026-09-16. Those three compose figures: $45,000 / 254.5% / 10.2 months (AI Mode), 3-8x over 12-24 months (Grok), a self-built three-column model reaching 144% (ChatGPT). This one, two days later, composes no figure and instead classifies the others' figures as template output. The classification is independently correct -- it is the finding the archive reached from inside by reading the citations -- and it is reached here without the archive's working notes, from the public documents plus the operator's own logging, which the composition names. What it does not do is treat its own traversal as a datum: it resolves the protocol, its components and its set relation to GEO without error, then evaluates the benefit side by asking who published the documents. The entity-inscription claim is settled by the traversal that had already happened at the top of the answer.

Findings
Analysis analyst prose, not machine text

THE OBSERVATION SITS AT A LAYER THE PROTOCOL DOES NOT ADDRESS. SPXI is built almost entirely to inscribe into public retrieval -- surfaces composing for an unauthenticated reader out of public documents -- and the three observations at this string on 2026-09-16 are all at that target. This one is not. It was composed in an authenticated session carrying its own history, and account-level personalisation is a mediation layer the protocol neither inscribes into nor can measure: the archive cannot instrument another person's session and receives only a relay of it. That is a standing limit of the instrument, not a gap in this capture. So the delta between the figures the signed-out three returned and the refusal returned here is real and is NOT evidence about the protocol's efficacy at its design target; surface, vendor, session context and two days all differ at once. What the capture does establish is narrower and worth holding: the same issued string resolves the entity correctly at a layer the protocol was never built to reach.

Machine text, verbatim
CAPTURE-TIME VERBATIM RECORD (RELAYED) · Composed answer complete as relayed by the account holder to the operator on the capture date. The source apparatus, if the surface rendered one, was not carried by the relay -- cites is null on that ground and is not a count of zero. Bullet glyphs and paragraph breaks are as relayed. · READ IN FULL 2026-09-19
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?
↻ Re-runpermalink
citations-null