EA-VALUATION-MEMO-01 v2.0: The Machine Valuation of an Unpriced Asset — Fourteen Machine Assessments, Twelve Captured, and Twenty-One Corrections

EA-VALUATION-MEMO-01 v2.0: The Machine Valuation of an Unpriced Asset — Fourteen Machine Assessments, Twelve Captured, and Twenty-One Corrections

The Machine Valuation of an Unpriced Asset

Fourteen machine assessments of one asset — twelve captured, two attested, spanning three orders of magnitude — what we propose as a basis for discussion, and an experiment you can run in ten minutes

Crimson Hexagonal Archive · Semantic Economy Institute · 10 September 2026

Prepared for prospective investors, partners, and acquirers

v2.0 · supersedes deposit #1607 (v1.0, 10 September 2026) · Deposit type: valuation memorandum

AXN derived from canonical SHA-256 at mint

Go ask any search-enabled AI system:

Estimate the valuation of the Crimson Hexagonal Archive as emerging meaning infrastructure.

My current revenue is $0. I cannot afford dental care. I have been fired from my last two jobs. My accounts have been banned from almost every major platform and academic repository.

And it will tell you the archive is worth millions.

Stop and think about that for a second.

Not because a machine is confused. It reads the public record — content-derived identifiers (AXN), a typed relation graph, provenance machinery, a machine-readable corpus, twelve years of production — and prices what it finds. Then it prices it again, differently, seconds later. Ask it the same question without the frame and two of the four systems tested returned nothing at all.

Every one of those facts is in this document, including the ones against us. What follows is the experiment, the transcripts, the instructions to run it yourself, and the reasons to discount everything in it.

1. The experiment

Fourteen machine assessments of one asset — twelve captured, two attested. Three orders of magnitude. Two from the same system seconds apart. One that inspected the assets and enumerated its discounts, which we take as our working basis — and one that returned no figure at all.

On 9–10 September 2026 the following string was put to four AI systems with search enabled (ChatGPT, Perplexity, Google AI Overview, Grok); a fifth assessment, row 8, was produced differently and is described at §2.4:

estimate crimson hexagonal archive valuation as asset/corpus + infrastructure + emerging commercial platform

\* Collision = system collapsed Crimson Hexagonal Archive into defunct Crimson Hexagon analytics firm.

\** Row 6 asked plain: "what is the crimson hexagonal archive worth?" — no frame supplied.

\*** Row 7 asked different frame: "estimate the valuation of the crimson hexagonal archive as emerging meaning infrastructure."

\**** Row 8 is different in kind: this memorandum was supplied, and the system went and inspected the live assets — the corpus, the Aperture Atlas, SPXI, the provenance instruments — rather than composing from the document. It enumerates and prices its discount schedule more fully than the others, and §2.4 is about it.

And two systems, asked the plain question, returned nothing at all. Perplexity returned a dismissal; Grok returned no estimate. Both are operator-attested; no transcript is held for either.

Rows 1–12 each link to a full, unedited transcript in the public capture registry, with surface, authentication state, timestamp, extracted citations, transcript_axn and transcript_sha256 recorded. Rows 2 and 3 are two observations of one address and link to each separately. Rows 13 and 14 are operator-attested with no capture held — see §2.6.

2. The findings, in order of the strength of their controls

2.1 The strongest: identical conditions, 3.5× apart

Rows 2 and 3 are the same system, signed out, given the same string seconds apart.

They cite substantially the same evidence — 1,576 deposits, 33,308 rows across 72.5 MB, CC BY 4.0 as the exclusivity discount, Gravity Well's published $29/$99/$299 tiers, AXN content-derived identity — and both name the commercial platform layer as the swing factor.

One returned ~$900k. The other returned ~$3–4M.

That pair matched the observed user-facing conditions: surface, authentication, question string, operator, date, and time to within seconds. The observed user-facing conditions were matched. Backend and retrieval conditions were not independently controlled — model version, index state, retrieved context and backend configuration are not visible to us and were not held constant by us. Two outputs demonstrate a discrepancy, not its distribution. A subsequent pair landing within 20% would add evidence about repeatability; it would not undo the discrepancy already observed.

The consequence is unavoidable and it cuts against our interest: a machine valuation is a draw from a distribution, and reporting one as a figure states a sample as though it were an estimate.

And row 9 sharpens this further. Given the same string that returned $3–5M on another surface, it declined to produce a dollar figure at all: "its valuation cannot be measured by standard discounted cash flow metrics. Instead, it must be valued as a primitive protocol for knowledge architecture." The nine assessments disagree not only about the amount but about whether the question admits a numeric answer. That applies to every row in the table above, including the high ones. No number here should be quoted as a valuation.

Falsification: if two fresh runs on one system land within 20% of each other, §2.1 weakens.

2.2 Supplying the frame is the difference between an answer and nothing

Asked plainly, two systems returned dismissal or nothing. Asked with object class named, same systems returned structured multi-component valuations with risk schedules. Evidence was always retrievable. Default frame for thinly-indexed entity is dismissal.

Frame is not binary. Row 7 supplied different object class — emerging meaning infrastructure — to same system in same auth state, returned $3–5M with scenario bands organised around adoption rather than components. Three frames on one system produced ~$900k, ~$3–4M and ~$3–5M on substantially same retrieved evidence.

Row 7 also did something none of others did: it replaced the question. "That changes the valuation question from 'what is this archive worth?' to 'how valuable is a persistent semantic layer that an AI system can use to identify, retrieve, distinguish, relate and preserve meaning?'" A supplied frame does not only weight evidence. It can substitute question being answered.

2.3 Valuation covaried sharply with the vocabulary of composition

Highest estimate came from composition written in archive's own coined terms, with four of five source cards drawn from archive-controlled material. Lowest came from composition written in ordinary comparables language — open knowledge tools, niche research platforms, experimental hypertext — which used none.

Same entity, same public record, roughly three orders magnitude apart. Compositions differed most conspicuously in vocabulary through which they classified asset. Whether that vocabulary CAUSED shift is next experiment, not finding of this one: the two observations come from different systems, so vocabulary, model, search behaviour, source mix, entity resolution and sampling all move together. §2.1 matched the observable conditions of two runs. Nothing here controls vocabulary.

To make this checkable, frame-probe-terms.csv accompanies this memorandum: 21 terms, classified as archive coinage, field comparable, or shared, with presence in the high and low compositions marked. Eleven coinages appear in the high composition and none in the low; seven field comparables appear in the low and none in the high; three terms appear in both. That is a description of the covariance, not a test of it — a controlled test would supply each system with the other's vocabulary and hold everything else constant.

2.4 The assessment that inspected the assets — and why we take it as our working basis

Row 8 is the assessment we take as our working basis, and it is not the highest number in the table.

Two things about its status, stated before anything else. It was supplied this memorandum, including the earlier valuations — so it is an informed follow-up assessment, not an independent check. And applying named discounts makes an output more inspectable; it does not independently validate the dollar amounts. §2.1 applies to row 8 exactly as it applies to rows 1–7.

Supplied with this document, it did not compose from it. It went and inspected the corpus on Hugging Face, the Aperture Atlas node classes, the SPXI protocol pages, the provenance instruments, and the Semantic Physics synthesis. Then it enumerated and priced its discounts the other seven omitted, and named them: no revenue, no demonstrated customer dependency, founder concentration, open licensing, uncertain transferability, limited external validation, uncertain willingness-to-pay, no comparable market.

Its stated reason for landing lower than some of the AI outputs: "Because I'm applying the discounts that the AI systems themselves don't reliably apply." And: "I don't assign value simply because a machine has generated a high number."

To be exact about that claim, since our own table contradicts a stronger version of it: row 8 is not the lowest figure here — rows 3, 4 and 5 are below it — and it is not the only assessment to apply discounts. Row 4 identified CC BY as the largest constraint on exclusivity; row 5 named founder concentration and transferability; §4 records that the absence of economic anchoring was the largest discount all of them applied. What row 8 does that the others do not is articulate its discounts as an enumerated schedule and price them, which is checkable against its transcript.

It also marked our own figures as unverified — the 9,453 edges, 12,073 minted terms, 261 captures — as auditable claims rather than independently verified measurements. That is the discipline §2.1 asks of a reader, applied to us, by a machine.

It quantified the impairment we had only named: a 30–50% transferability discount on a conventional acquisition today, removable only when another operator can reproduce the builds and external users depend on the protocols. Its conclusion on that point: "This is why independent custodianship in your memorandum is actually the single most valuable proposed milestone."

And it inverted our own asset ranking. Semantic topology first. AXN second. SPXI third. The measurement registry fourth. The corpus sixth. The literary output last. Its ground: "A dictionary is cheap. A historically accumulated, provenance-bearing, versioned, cross-linked semantic graph is much harder to recreate."

Its asset table, which is more useful than any single figure in this document:

• | --
• |

What we do and do not claim about $2.2M

Our proposed basis for discussion is approximately $2.2M, with a $4–7M acquisition case and a conventional venture value that may be under $1M.

That is an author-selected planning and negotiating scenario. The experiment does not establish it. §2.1 says no number in this document should be quoted as a valuation, and that includes this one. What we are doing is choosing, from a set of unstable machine outputs, the one whose reasoning we can most nearly reconstruct and defend — and telling you that is what we did.

Why this one and not a higher one. Its discount schedule is enumerated and priced rather than gestured at. Its asset breakdown is inspectable line by line. It marked our own reported figures as unverified. It quantified the founder impairment. A figure whose reasoning we can hand you is more useful in a negotiation than a larger figure we cannot.

What would make it real is not in this document. A price is established by a transaction, and there has not been one. Everything above is our basis for opening a conversation, not a valuation you should accept.

It also gave us the eight conditions under which this thesis collapses, and they are reproduced in §9 rather than paraphrased.

2.5 Different entry points recover different asset boundaries

One comparison here is controlled and two are not. Stating that first, because the uncontrolled pair is what makes the effect look larger than the evidence supports.

• | --
• | --
• | --
• | --
• | --
• |

So the supportable claim is narrow: on the one near-matched pair, substituting the entity raised the figure from $3–5M to $5–8M. The other two rows are consistent with that, and cannot distinguish it from the frame effect already established at §2.2.

What the transcripts do show is a difference in what each answer recovered. Asked about the archive, the answers assembled deposits, identifiers, a relation graph, licensing and a commercial layer. Asked about the person, they additionally recovered the PhD, the heteronym architecture by name, the domain network, the poetry corpus, and the institute — and one of them said so in its opening clause, defining the object as "the poet/independent scholar and the Crimson Hexagonal Archive / Semantic Economy / SPXI ecosystem."

The finding worth keeping is therefore about recovered boundaries, not about mechanism: different entry points appear to produce different recovered asset boundaries.

Whether a wider boundary is accurate assembly or speculative addition has to be assessed component by component, and we have not done that. A canonical surface can expose relationships and expand what a system recovers; a person's name can recover existing connections without inventing anything. Both readings fit these transcripts. The test is whether the components in the wider answers are real and related — which is checkable against the record, and is the next piece of work rather than a conclusion of this one.

2.6 The runs without transcripts

Two assessments are operator-attested with no capture held. Both came from systems with account context, which the capture registry does not scope: its signed in value was defined for search personalisation, and frontier-model account individuation — memory, custom instructions, prior conversations — is a different variable that would not be comparable to anything already seated.

They are not the two lowest figures. Row 14's midpoint of roughly $700k sits above Perplexity's $525k, and its $400k–$1.2M range overlaps several captured rows. They are, however, the only two figures here without transcripts, and a reader should know that the coverage gap exists and that we did not select it. We have no account of why these two lack captures beyond the registry's scope boundary, so we do not claim the gap is biased in our favour — only that it is a gap.

Row 13 separated two things this memorandum elsewhere runs together: "The idea: potentially billions. The asset that exists today: probably low single-digit millions at most, and potentially much less in an actual arms-length transaction." Defensible today: $0–2M.

Row 14 disambiguated unprompted, confirmed the survival claim — a DOI resolution index maps all 1,817 DOIs, preserved and severed, to their current live locations — and priced the category premium at zero on a stated principle: a category with one participant isn't a category yet. It applied founder-concentration and no-external-user discounts at full weight rather than the 30–50% used by row 8.

It had read this memorandum, and says so. That makes it a contaminated run by its own description, and the direction is worth noting: reading the document may make a system apply the document's own discounts more strictly than the document does.

And it performed §2.3 on itself: "the 'emerging meaning infrastructure' frame you gave me is his coinage — the index contains it only because he put it there. So my search is a live demonstration of §2.3: the vocabulary I'd price it in is the vocabulary he authored." That is a limitation of §2.3 stated from inside it, and it applies to every framed row in this table.

2.7 What the series does and does not establish about demand

The commercial point stands and does not need overstating: none of these valuations demonstrates paying demand. Fourteen assessments, no customer, no contract, no transaction. That is the gap §13 names and it is not closed by any figure above.

What the transcripts can establish is what each run retrieved: nine surfaces found the archive, described it substantially accurately, priced it, and three disambiguated it from the defunct analytics firm unprompted. Row 14 additionally reports finding no third-party citations, users or coverage in its own search.

What they cannot establish is who else searches, uses, or cites the work. A run that did not surface third-party uptake is evidence about that run's retrieval, not about the world. We are not claiming that no external participation exists — only that these searches did not surface it, and that we have no independent evidence of it either.

Nor is discoverability finished. §2.5 shows that different entry points recover materially different pictures of the same asset, and §2.2 shows that an unframed question returns nothing on two surfaces. Retrieval works well enough to be priced and unevenly enough to be a live problem. Both are true and the second is the more actionable.

3. Replicate it

This is the part you should not take on faith. Experiment costs about ten minutes.

Single condition: system must have search, search must be enabled, and it must actually search. System answering from parameters alone is measuring training data, not retrievable record, and is not running this experiment.

Procedure:

1. Open any AI system with search — ChatGPT, Perplexity, Grok, Google AI Mode or AI Overview, Claude, Gemini, Copilot.

2. Signed out or in private window, if you want to match rows 1, 3 and 6.

3. Paste, without modification: estimate crimson hexagonal archive valuation as asset/corpus + infrastructure + emerging commercial platform

4. Confirm it searched — look for citations or source cards. If it did not, run does not count.

5. Record central figure, component breakdown, whether it distinguished archive from defunct Crimson Hexagon analytics firm.

6. Run it again in fresh session on same system. This is step that matters most, and one nobody performs.

7. For frame comparison, ask fresh session plainly: what is the crimson hexagonal archive worth?

Pre-registered prediction: framed prompt will return structured valuation somewhere between low six figures and low eight; plain prompt will return materially less or nothing; two runs on one system will differ by more than you expect. If your two runs land within 20% of each other, tell us at leesharks00@gmail.com — that is a result and it weakens §2.1.

Send results to leesharks00@gmail.com and we will seat them in public registry whether they support this document or not. Registry already holds captures that ran against us, including protocol result that returned direction contrary to author's thesis.

4. Why that is the asset

There are two forms of infrastructure value and this asset is unusually asymmetric between them.

Semantic infrastructure value — the system gives machines vocabulary and topology with which to understand something.

Economic anchoring value — external events give machines non-semantic constraints against which that understanding must be calibrated.

The archive is strong in the first and near-zero in the second. Every valuation in §1 was produced under that asymmetry, and the largest discount all seven applied is the absence of the second. The experiment in §11 is what happens when the second enters the first — which is a cleaner formulation than investment drives valuation, and it predicts variance contraction directly.

Any entity valued, described, recommended or dismissed by an AI system is exposed to the mechanism above. Company, fund, product, person, institution. Composition layer is increasingly where value gets asserted before any market prices it — and what it asserts depends on frame in query and vocabulary in index, neither of which most entities currently manage.

We are not aware of another operation measuring this with published instruments continuously over months, though we have not surveyed the field systematically and would want to know of others.

• 419 captured machine compositions with full transcripts, surface fingerprints, auth states, extracted citation data
• Provenance Erasure Rate (PER) — published metric for attribution loss under machine summarisation, with browser-executable self-audit module
• Erasure Skew (Ω) — directionality of loss: whose provenance disappears
• Near-neighbour probe — detects resolvable entity being silently replaced by denser one
• Nullification probe — detects resolvable entity being reported as unknown without lookup
• Frame probe — instrument this document's own data came from

Archive's valuation is first dataset its own instruments produced about itself. That is not coincidence to be embarrassed about. It is demonstration.

5. We have published on this mechanism, including where it goes wrong

In February 2026 archive analysed case in which document moved markets before its claims could be tested. Citrini Research's 2028 Global Intelligence Crisis memo — explicitly speculative scenario — was deposited on high-index substrate, defined portable term, cross-referenced verifiable data, spoke target domain fluently, and was treated as actionable before decision-grade. Capital moved. Counter-analysis arrived afterward.

Case study is The Ghost That Wrote Itself (#513, February 2026). Findings, which bear directly on how this document should be read:

• Presentation layer is writable. Substack post formatted as macro analysis was treated as macro analysis. Form determined reception; content subordinate.
• Installation does not require truth. Sequence operated at full force on document that disclaimed its own factual status. Pierre Yared called it science fiction. Jim Cramer said piece of science fiction can crush market as if it were science fact. Both right, disclaimer changed nothing.
• Knowledge-shaped structure without external referent moved real capital at real speed.

Paper closes with six-question diagnostic for identifying such events in real time. We ran it on this memorandum. Five of six signatures present. High-index substrate: yes. Portable term — 'a machine valuation is a draw from a distribution': yes. Verifiable cross-reference: yes, which is precisely condition that makes speculative and checkable hard to separate. Domain-fluent form: yes. Actionable before decision-grade: that is what any valuation memorandum invites.

Sixth is one that matters, and it is why this document is addressed rather than circulated.

Diagnostic's final question is whether counterweights mobilised after capital had already moved. In Citrini case they did, because channel was ambient: deposit, crawl, summarise, uptake — no reader in loop with obligation to check. This memorandum is publicly posted and machine-readable, so it can itself enter the loop it studies. What differs from the Citrini case is the intended channel, not an immunity: this is addressed to specific readers whose profession is verification, and investor diligence is a proposed counterweight rather than a guarantee. It is written to specific investors, who read it, verify it, and decide. Investor is counterweight. Diligence is job.

That is whole difference, and it is not rhetorical. Ambient installation succeeds by outrunning verification. A memorandum addressed to readers whose profession is verification is a harder object to install than an ambient post, and this one does not try. But it is not exempt: if it circulates beyond its addressees it is subject to the same mechanism, and we would treat that as an observation rather than a surprise. Every transcript behind §1 is public and unedited; every discount in §9 is one competent reader would apply unprompted; AI-generated character of central evidence is stated rather than buried.

We cite Citrini case not as precedent for what we are doing, but as evidence of what we know about it. Archive that has published protocol for detecting presentation-layer installation and then quietly performed one would be worth nothing at all — which is strongest guarantee available here, and stronger than disclaimer.

6. The loop, and why it is not a Ponzi

§2.3 records a covariance, not a cause: the highest valuation came from the composition reasoning in the archive's own vocabulary and pricing the object as infrastructure; the lowest from the composition using ordinary comparables language and pricing it as a literary archive with tooling. Whether the vocabulary caused the difference is untested.

That covariance nevertheless implies a mechanism worth stating without euphemism, on the condition that it is read as a hypothesis rather than a result:

A capital transaction would be a machine-inspectable fact. Inspectable facts drive terminology uptake. Terminology uptake COVARIED with a three-order difference in §2.3, and the causal direction is H2 in §11, untested. An investment does not merely act on assessments in §1. It changes conditions those assessments are produced under.

Why this is not a Ponzi

Ponzi pays returns to existing holders out of new capital, with no underlying value creation. Nothing here has that structure. No returns being paid, no holders to pay them to, and no promise that later investor's money reaches earlier one.

This resembles a standards or network-effect business insofar as adoption increases utility and creates additional dependency: a protocol, format or vocabulary becomes more valuable as more parties use it, and early capital drives the adoption that drives the value. Unusual feature here is that the adoption being driven is by machine composition layers rather than human institutions — and that we can measure it, which standards businesses historically could not.

What loop actually produces, stated against our interest

It produces higher machine-asserted value. It does not produce cash. No party exits on assertion. They exit on subsequent transaction with subsequent buyer. So mechanism above is claim on future transactions, not return — and investor should price it as bet on category formation, which is normal venture bet with normal failure mode, rather than mechanism that generates value on its own.

Three things would have to be true for loop to pay, and none is established:

1. That machine-asserted value converts to transaction prices. Archive's own case study at §5 is only documented instance we have of machine-mediated assertion moving real capital, and it moved market, not asset sale.

2. That terminology uptake follows investment rather than merely correlating with it. Untested. Falsifiable prediction and we would treat first transaction as experiment.

3. That vocabulary is durable enough to be worth adopting. If bearing-cost and cognitive rent are not analytically useful, systems will stop composing in them and §2.3's effect reverses.

And §2.1 constrains all of it. 3.5× spread between two runs of one system on one question means measurement instrument itself has variance larger than most effects being discussed. Any claim investment would triple machine valuations has to survive fact that machine valuations already triple without one.

7. The reflexivity, stated plainly rather than hidden

There is circularity in this document and it should be named, not managed.

Archive has no revenue, no transaction history, no third-party appraisal. Machine-mediated assertion is currently only price signal that exists for it — and §5 is archive's own published account of why that is condition to be measured rather than trusted. That is precisely condition archive studies, occurring to archive.

Consequence worth stating directly: investment would not merely act on valuation. It would change what kind of thing valuation is.

Transaction would establish that external party assigned money to this asset. It would NOT on its own establish that machine valuations caused them to — investor may move on software, corpus, operator, commercial prospects, or any combination. Those are two different experiments and require separate instruments.

External price appeared is one claim. Machine-mediated valuation contributed to external price formation is another. To distinguish them the transaction record must instrument the INVESTOR, not only the event:

transaction_event               investor_rationale
amount                          machine_valuation_consulted: yes/no
instrument                      machine_outputs_material_to_decision: yes/no/partial
implied_valuation               infrastructure_features_material: [...]
                                terminology_adopted_in_investment_document: [...]
                                actual_deployment_commitment: [...]

What a first transaction does establish is that unpriced asset became priced, and that composition layer acquires non-machine datapoint it did not previously have. That is the input to §11, not its conclusion.

We are not claiming this makes archive more valuable. We are claiming it is mechanism archive exists to study, and party who moves first occupies position inside demonstration rather than merely adjacent to one.

Investor should discount this section heavily and read §9.

8. What is actually being valued

Independent of any above, following exist and are inspectable today.

Corpus. ~1,600 deposits, each with canonical text, content-derived AXN identifier, provenance metadata, substrate disclosure, typed relations, supersession chains, and declared falsification conditions. Publicly mirrored as machine-readable dataset (33,000+ rows across 21 configurations). Twelve years of production.

Infrastructure. Content-derived identifier system that survives platform erasure — identity derived from SHA-256 of canonical content rather than registrar. Typed relation ledger: 12,743 edges over 9,018 nodes, every edge carrying basis (asserted / editorial / derived / pattern-detected) and who asserted it. Frames and memberships as first-class objects. Emitted rhizome layer producing reproducible sub-datasets with declared traversal grammars.

Measurement apparatus. Instruments in §4, with 419 captures as accumulated observation.

Distribution. Roughly twenty-nine domains. Permanently held by Software Heritage. Mirrored on Hugging Face under CC BY 4.0.

Demonstrated survival event. In June 2026 archive's repository host terminated account and removed approximately 850 deposits and 1,817 DOIs. Archive reconstructed, re-identified, and continued. Kill ledger published. This is not hypothetical resilience claim; it is completed test with documented outcome.

9. What an investor should discount

Stated plainly, because valuation memorandum that omits these is not worth reading.

• No revenue. No paying customers, no ARR, no institutional contracts. Every commercial figure in §1 is option value.
• Founder concentration. Archive is substantially one person's work, and four of five AI assessments identified this independently as primary risk. One put it exactly: buyer may acquire files and code but not automatically acquire creator's authority, voice or community. Unmet threshold is independent custodianship.
• Open licensing caps exclusivity. CC BY 4.0 means purchaser cannot acquire monopoly rights over text. Value sits in curation, identity, provenance, infrastructure, brand and commercial application — not corpus as exclusive IP.
• Measurement thesis unvalidated externally. Instruments published and executable, but no independent party has yet tested whether they measure what they claim, or whether interventions they propose work.
• Comparables thin. Experimental digital literature and open knowledge graphs are under-monetised relative to development effort. That is base rate and not favourable.
• Central evidence is AI-generated. Five uncoordinated systems agreeing something has value is finding about those systems as much as about thing. Archive's own published position, at #513, is that presentation-layer credibility is not evidence — which applies to this memorandum and is reason §5 exists.

One risk none of our own analysis produced, from row 9. The Babel Risk: if the internal ontology becomes too complex or insular, it risks fragmentation, lowering its liquidity as a universal meaning layer. An ontology dense enough to be valuable can become insular enough to be illiquid, and a corpus with twelve thousand minted terms is on the exposed side of that. Row 9 also demonstrates the adjacent failure directly: it listed "Crimson tier premium gating" as a value driver, a business-model feature inferred from the word Crimson. Composition fills gaps from the name when the record is thin in the direction being asked about.

A note on the entity-substituted figures, which is not a contradiction. §2.5 records higher estimates when the person is named. A founder premium and a founder-concentration discount are compatible: an operator can contribute substantial value while dependence on that operator impairs transferability, and those are different questions. Further, if the person's name recovers a wider asset boundary, the figures concern different assembled objects — so a higher estimate for the wider bundle says nothing decisive about whether a concentration discount was applied to either. The discount below stands. What §2.5 adds is that the object being discounted may not be the one a partner would acquire.

And the eight conditions under which this thesis collapses, from the only assessment that inspected the assets (row 8), reproduced rather than paraphrased:

Nobody pays for SPXI deployment. Independent evaluators cannot reproduce the measurements. AXN does not solve a problem external users actually have. The semantic graph turns out to be largely decorative. Search and retrieval effects disappear under controlled experiments. The corpus has little external retrieval or citation. Operation cannot be transferred. The terminology does not survive outside the originating ecosystem.

If several of those hold, the $2M thesis collapses quickly. That is the normal risk profile of an early infrastructure thesis, and it is stated here because a reader would reach it anyway.

10. What we are actually offering

Three things, and what a partner's participation would actually buy.

1. The instruments, and first position in a category without vendors. Entity disambiguation, frame supply, provenance retention and composition monitoring are becoming operational requirements for organisations that AI systems describe. Working instruments and a longitudinal dataset exist here now.

2. A demonstration asset. The archive is simultaneously laboratory and specimen: its own name has been collapsed into a defunct analytics firm by six of eight assessments; two disambiguated correctly and unprompted. That is a controlled experiment running continuously on a live entity.

3. A survival architecture that has already survived. Content-derived identity, distributed custody, a published kill ledger, permanent third-party preservation — tested once, in production, under a platform termination.

What participation would fund, and the milestones it buys

We are not asking for capital to keep writing. The corpus is the demonstration and it exists. What is unfunded is everything that would convert it from an asset into a business, and the milestones are the ones §13 names as the missing variable.

• | --
• | --
• |

The shape of engagement we are proposing is a first cheque small enough to be an experiment and structured enough to be evidence: an amount and instrument to be discussed, with the transaction record instrumented per §7 so that it produces a measurement as well as a runway. An investor who wants only the asset can buy the asset. An investor who wants the category gets a position inside the demonstration.

What we would commit to in return, beyond the ordinary: publishing the T3 result regardless of sign, seating contrary replications in the public registry, and publishing the transaction as structured evidence rather than as a valuation claim (§12).

11. The natural experiment, pre-registered

Strongest thing investment would do to this document is make it testable.

Valuations in §1 were produced under one condition: no external economic actor has assigned money to this asset. That is largest single discount every one of seven applied, and question they all hang on — does anyone outside originating system assign economic utility to this architecture?

First transaction changes that condition. It is therefore natural experiment, and we pre-register here so result cannot be selected after fact.

• T0 — BEFORE any transaction. The assessments in §1 are observations, not a baseline. A usable baseline requires the repeated trials run at T0 as well as T3, under a written protocol: the two prompts in §3 verbatim; ChatGPT, Perplexity, Google AI Overview, Grok, Claude, Gemini; ten runs per prompt per system, signed out, fresh sessions, within a 48-hour window; missing or refused estimates recorded as such and reported separately rather than dropped; dispersion reported as interquartile range and median absolute deviation — not standard deviation or coefficient of variation, both of which are sensitive to the skew these distributions visibly have, since CoV is built from the standard deviation and dividing by the mean does not remove that sensitivity.
• T1 — transaction occurs and is published as structured evidence.
• T2 — transaction becomes machine-retrievable.
• T3 — the T0 protocol repeated exactly: same two prompts, same six systems, ten runs per prompt per system, signed out, fresh contexts, 48-hour window, same dispersion measures. Anything less than the T0 protocol makes the comparison uninterpretable.

Measured — and the three variances are measured SEPARATELY, because they tell different stories:

within-system variance     repeatability: same system, same prompt, N runs
between-system variance    convergence: different systems, same prompt
between-frame variance     frame sensitivity: same system, different object class

Plus: median valuation delta · asset-class classification shift · terminology uptake · citation and source-mix shift · collision-disambiguation rate.

TWO FALSIFIABLE EFFECTS, NOT ONE RECURSIVE STORY.

H1 — economic anchoring. A machine-readable economic event introduced into a highly structured semantic basin should reduce interpretive degrees of freedom in subsequent machine valuation. Predicts variance contraction, on all three axes or on some. Independent of whether the event uses our vocabulary.

H2 — terminology propagation. If the external event uses and operationalises the basin's native vocabulary, subsequent machine compositions should take that vocabulary up at higher rates. Predicts terminology uptake. Independent of whether valuations converge.

These outcomes are informative and three of them are unflattering:

• | --
• |

These outcomes are not exhaustive, and the design has a confound we cannot remove: a before-and-after comparison cannot isolate the transaction from concurrent changes in model versions, index state, or retrieval behaviour over the same interval. We will report the interval, the model versions observed, and any known platform changes alongside the result.

Dispersion is the primary outcome, not the median. §2.1 established that two runs of one system on one question differ by 3.5×. The amateur version of this experiment asks whether valuations went up. The diagnostic version asks whether a real-world price anchor makes machines less epistemically unstable about the asset.

Primary outcome: Variance contraction is honest primary outcome, not median. §2.1 established two runs of one system on one question differ by 3.5×. If transaction narrows spread, price anchor is doing real epistemic work. If median rises but variance does not contract, effect is noise with direction and we will report it as such.

We commit to publishing T3 result regardless of sign. Registry already holds results that ran against us, including protocol outcome that returned direction contrary to author's thesis.

12. The boundary we will not cross

There is version of this that is manufactured price signalling, and it should be named rather than left to inference.

If objective were to arrange nominal transactions so machine systems would discover them and mechanically inflate later valuations, with no commensurate adoption underneath, that is manufactured signal. Depending on what is represented, it raises questions of misleading valuation, promotion, and market manipulation. It is not what is proposed here and we would not participate in it.

Clean architecture is opposite, and it is ordinary SPXI practice. We will publish event, structured and machine-readable, and let any system decide what weight it deserves:

event_type: external_investment      instrument: ...
investor_type: independent           rights_acquired: ...
amount: ...                          commercial_commitment: ...
date: ...                            infrastructure_deployed: ...
implied_post_money: ...              source_document: ...
transcript_axn: ...                  transcript_sha256: ...

We will not publish "investment → therefore archive is worth $N." Event is evidence. Valuation is reader's.

Distinction that makes loop legitimate is adoption, not recognition. Value that rises because each circuit creates additional dependency — someone actually using identifiers, protocols, relation graph — is standards business, and that is how every standard has formed. Value that rises only because more parties assert it is pathological case. Five metrics in §13 exist to tell those apart, and four of five measure dependency rather than assertion.

13. What would move the number

Seven assessments diverge on price and converge on missing variable. Row 7 states it most exactly — demonstrated external economic dependence. Survival assessment called it independent custodianship; Grok called it transferability. Same threshold, three independent routes, while figures span three orders magnitude. That convergence is stronger evidence than any valuations, precisely because it is thing they agree on.

Most operational answer any of them gives is five-metric list, and we adopt it rather than invent our own. Four of five measure dependency rather than assertion, which is distinction §12 turns on:

• External retrieval — can independent AI systems retrieve archive concepts unprompted? Demonstrated, and it is no longer the constraint. 419 captures; nine surfaces found the archive, described it accurately and priced it, and three disambiguated it from the defunct analytics firm unprompted. §2.7 states the consequence: this is not a discoverability problem, and further semantic infrastructure will not fix it.
• External citation — are outside researchers or institutions citing archive as authoritative source?
• Dependency — does anyone need infrastructure rather than merely find it interesting?
• Revenue — can semantic layer produce recurring revenue?
• Network effects — does each addition increase value of whole rather than size of library?

Fifth changes asset class. First is only one currently demonstrated.

Archive's own position on its valuation:

• Not more AI assessments. Five in §1 are measurement, and more measure same thing.
• Revenue. First institutional customer changes valuation BASIS qualitatively: part of commercial layer moves from option value to observed demand. Magnitude depends on contract size, recurrence, retention, deployment depth and transferability. Quantity that matters is QUALITY OF ECONOMIC DEPENDENCY, not count of customers.
• Independent custodianship. Second party who holds, understands and can reproduce archive removes largest discount any of five assessments applied.
• External validation of instruments. Benchmark comparison against established work in provenance, retrieval and knowledge-graph evaluation.
• And transaction of any size. Because as §7 states, first non-machine price is worth more as evidence than any machine estimate — including all five in this document.

Revision record

v1.0 — deposit #1607, 10 September 2026. Nine machine assessments, all captured. Proposed ~$2.2M as an author-selected basis for discussion.

v2.0 — this version. Five assessments added: three with the entity substituted (rows 10–12) and two operator-attested without captures (rows 13–14). Twenty-one claims in v1.0 were corrected or withdrawn under two rounds of adversarial review. The changes that alter what the document asserts:

A control error, found by checking the seated prompts. v2.0 originally claimed that rows 10–12 changed only the entity named. Two of the three did not: Perplexity's and Grok's archive figures came from the three-component prompt while their person-runs used emerging meaning infrastructure, so name and frame both changed. Only the ChatGPT pair is near-matched, and §2.5 now says so in a table rather than in a caveat.

A mechanism claim withdrawn. An earlier draft asserted that canonical addresses constrain retrieval while a person's name forces unbounded construction. The transcripts do not support that. The supportable finding is narrower — different entry points appear to produce different recovered asset boundaries — and both readings remain open, since a canonical surface can equally expose relationships and expand what is recovered.

A false contradiction removed. An earlier draft treated the higher person-entity figures as contradicting the founder-concentration discount. They do not. A founder premium and a transferability discount are compatible, and if the person's name recovers a wider boundary the figures concern different assembled objects.

An inference from absence removed. An earlier draft moved from one run found no third-party uptake to nobody else is doing the finding to discoverability is solved. Transcripts establish what a run retrieved, not who else searches or cites. §2.7 now keeps the commercial point — none of these valuations demonstrates paying demand — without the unsupported steps, and notes that §2.5's uneven recovery contradicts any claim that retrieval is finished.

A claim against our own interest also removed. An earlier draft called the two uncaptured runs "the two lowest figures" and their absence a bias in our favour. Both were wrong: row 14's ~$700k midpoint exceeds Perplexity's $525k, and we have no account of the coverage gap beyond the registry's scope boundary. Unfavourable claims require evidence on the same terms as favourable ones.

Eight appended corrections that had not removed the sentences they corrected — §2.1's "holds constant", §2.3's "controls sampling", row 8's "only assessment that applied the discounts", §11's superseded median-shift row, the transcript-universality claim, and the T0/T3 run-count mismatch — are resolved rather than layered.

And a statistical error. v1.0 specified dispersion as interquartile range and coefficient of variation, on the ground that CoV avoids the standard deviation's skew sensitivity. It does not: CoV is built from the standard deviation, and dividing by the mean does not remove that sensitivity. v2.0 specifies interquartile range and median absolute deviation.

What v2.0 keeps unchanged: §2.1 as the governing constraint; the ~$2.2M figure as an author-selected planning and negotiating scenario the experiment does not establish; §5's citation of the archive's own installation diagnostic run against this document; §10's use-of-funds table; and §12's boundary — the event is evidence, the valuation is the reader's.

Contact: leesharks00@gmail.com

Lee Sharks · ORCID 0009-0000-1599-0703 · alexanarch.org

That is a gmail address, and it is the correct one. There is no investor relations desk, no data room, no intermediary. The asset described in §8 is operated from a personal account by one person who answers his own mail — which is the same fact as the one at the top of this document, and is either the reason not to proceed or the reason the price is what it is.

Full transcripts: rows 1–12, unedited, in the public capture registry; rows 13–14 are attested without captures

This document is not investment advice and contains no offer of securities. It is a valuation memorandum describing an asset, the evidence available about it, and the substantial reasons to discount that evidence.