Winning the Race: America’s AI Action Plan (The White House, July 2025) has three pillars — innovation, infrastructure, international diplomacy and security. Four of Pillar I’s sections bear directly on verification. CHECKED
| Section | p. | What it asks for |
|---|---|---|
| Build an AI Evaluations Ecosystem | 10 | NIST/CAISI guidelines for Federal agencies to run their own evaluations; “support the development of the science of measuring and evaluating AI models”; a NIST AI Consortium to establish “new measurement science that will enable the identification of proven, scalable, and interoperable techniques and metrics” |
| Invest in AI Interpretability, Control, and Robustness | 9 | a DARPA-led programme, with CAISI and NSF, because “the inner workings of frontier AI systems are poorly understood” and technologists “often cannot explain why a model produced a specific output” |
| Invest in AI-Enabled Science | 8 | automated cloud-enabled labs; and a requirement that federally funded researchers “disclose non-proprietary, non-sensitive datasets that are used by AI models during the course of research and experimentation” |
| Build World-Class Scientific Datasets | 8 | NSTC recommendations on “minimum data quality standards” for scientific data used in model training |
Read together, the ask is coherent and it is the right ask: a measurement science for a layer that currently has conventions instead of measurements. The rest of this note is about the one word in it that is load-bearing and unfunded — interoperable.
A metric transfers between parties when the receiving party can reconstruct what the number is a number of. That is what a unit does. “Three” transfers because “three metres” carries the metre with it.
A verification claim does not do this, and the failure is structural rather than sloppy. WP-73 of this series states it in one line: a verification claim names an artifact, but what was verified is a triple.
File headers record the artifact. Lock files record the library, elsewhere, unlinked to any claim. Toolchain pins record the toolchain, elsewhere again. No convention in common use binds all three to the assertion that depends on all three — so a claim survives intact while any of its arguments is replaced, and reads exactly as it did when it was true.
This is not a Lean problem and not an artefact of formal methods. It is the general shape of software provenance, and it is exactly the shape an evaluation score has: a benchmark number names a model. The harness version, the prompt template, the decoding parameters, the scorer, and the dataset revision are recorded elsewhere if at all. An eval result published against a model and a harness that both move is a claim whose arguments were never written down.
The Plan requires disclosure of inputs (datasets, p. 8) and asks for outputs (evaluation scores, p. 10). It says nothing about recording the environment a result was produced in — which is the argument that decays without anyone editing anything. OPEN
WP-94 of this series censused five verification communities and found five non-interchangeable words for the same object — a step that is not finished, standing inside an otherwise complete argument. Lean writes sorry. Coq writes admit and Admitted. Proof-carrying code writes axiom and trust. An avionics assurance case writes assumption. The differences are recoverable from grammar, not doctrine, and no term spans them.
A consortium convened to identify interoperable techniques will find that the communities it is convening cannot state, in shared vocabulary, where their arguments stop.
And it is open precisely because the layer that would need it — a proof- or evidence-exchange layer between parties — does not exist either. The window is real and it is not permanent. OPEN
The third gap is the one that costs the most and reads as the most technical. A gate that enumerates forbidden results cannot see a result nobody has thought of yet; a gate that enumerates permitted results closes the class permanently. WP-73 §6 gives the specimen: a theorem resting on an axiom emitted by compiled-code evaluation passed a step named “Kernel axiom check”, because the check was a text search for the one failure its author had in mind.
The same section gives the mirror failure, which is stranger. A checker counting only the “depends on axioms” form of a report was blind to the other form — a proof resting on nothing whatever, the strongest available outcome — and failed a file over its own three best results. A verification tool has two ways to be wrong, and the literature attends almost entirely to one of them.
A gate is a claim about the artifact it guards and is subject to the same rule as any other claim: it must be checkable on a fixture. A pipeline buried in a YAML file cannot be run on one. OPEN
This series did not build these in response to the Plan; they were built because the corpus needed them, and the Plan happens to describe the hole they fill. Four pieces, all in TOTOGT/geometry, all runnable by a third party:
| Piece | What it does |
|---|---|
| CS/verify-stamp | binds a claim to the triple and fails when they come apart, distinguishing MISMATCH (wrong artifact), STALE (environment moved), FAIL (real regression) — three outcomes that are identical to every automated reader today |
| the three tiers | written / sorry-free / kernel-audited, reported separately and never collapsed. Only the third is evidence |
| tools/axiom_gate.py | an allowlist, not a forbidden list; reads both output forms; rejoins the pretty-printer’s wrapped lines; and has its own fixture suite including the CI run that motivated it |
| the defect taxonomy | MISMATCH, STALE, FAIL, FALSE, VACUOUS, UNTRUSTED, MISATTRIBUTED (WP-73), and OVER-GENERALISED (WP-97) — each with what detects it, including the two that nothing detects |
The design rule under all of them is one sentence: a published figure must be produced by an artefact on disk, not typed. Tier 1 in this corpus is read out of the axioms.txt files the gates write; a number with no report behind it does not appear.
Recommending a tool on its successes is worth little. Here is what it caught in this corpus between 1 and 5 September 2026, all four defects belonging to its own author. RECORDED
| Found | What it was | Cost had it stood |
|---|---|---|
| UNDERCOUNT | the kernel-audited tier read 32 against a true 133. A glob swept two hardcoded directory shapes and missed a report holding 82 audited declarations — the largest in the corpus | a four-fold understatement in the one tier whose design premise is that its number cannot be typed |
| NO EVIDENCE | the audit tool wrote its probe to a temporary file, read two counts off it, and deleted it. A full-corpus overnight run printed to a terminal and left nothing | the repository could not distinguish a file audited last night from one never checked at all |
| FORBIDDEN LIST | a text search for one failure mode, reintroduced in three separate places by an author who had documented why it is wrong | a declaration the kernel never checked, counted as kernel-checked |
| OVER-GENERALISED | a published claim — kernel-verified, zero admitted steps — whose sentence was wider than its theorem, because a parameter fixed in the hypotheses read in prose as a constant of nature (WP-97) | a true theorem carrying a false sentence, indefinitely, with a green check beside it |
Corpus state as this note is written, every figure read from reports on disk by wp98-verify.py: 148 distinct kernel-audited declarations across five gate reports; 0 resting on anything outside the permitted three; 0 report lines the gate could not parse; 11 proofs resting on no axiom at all. MEASURED
That last figure was first written here as 8. Eight is the geometry-only count from the per-file ledger, carried into a corpus sentence without its scope — the same defect as WP-97’s, one week later, by the same author. wp98-verify.py refused the paper and printed the right number. That is the argument of this note arriving on the note, and it is left in rather than quietly corrected.
The Plan is written as a race. Its first sentence is that the United States is in one; its stated objective, quoted from the President, is “unquestioned and unchallenged global technological dominance”; its introduction closes “the AI race is America’s to win.” CHECKED A race is a zero-sum frame, and a zero-sum frame has a property that is easy to miss: a position won in it must be re-won. Dominance is not a holding. It is a lead, and a lead is a flow.
WP-38 of this series is about exactly this distinction and supplies the parameter that decides it. Two players: J, who holds a hub, has private information, pays a fixed carry, and moves strength per unit; V, who has speed, public information only, convex costs, and capped capture. The pivotal object is not who is ahead.
The hub is contestable if V has an action that transfers ownership with positive probability; non-contestable if that probability is zero for every action available to V. The preemption hazard λ is that probability; non-contestability is λ = 0.
Under non-contestability, no expenditure by V — of any magnitude, at any speed — transfers the hub. V may compete for flow rents; V cannot compete for position.
The Plan casts the United States as V throughout, and does so deliberately: speed, build rate, removal of friction, public strategy. The question WP-38 makes askable is not whether that is the right posture. It is which of the assets in the Plan have λ near zero — because only those are held rather than led.
| Asset the Plan pursues | λ | Why |
|---|---|---|
| A frontier model lead | high | Weights diffuse, capabilities are distilled, researchers move. A lead is preempted on the timescale of a release cycle. The Plan raises this λ itself and knowingly: p. 4 promotes open-source and open-weight models on the grounds that they “could become global standards” and therefore “have geostrategic value”. That is a bid for position through an instrument that makes position transferable. |
| Compute and infrastructure | middling | Fabs and grid are years of carry, which is real friction and raises the crossing threshold. But they are transferable in principle and in practice — Pillar III is largely about plugging export-control loopholes, which is an admission that λ > 0. A moat that has to be patrolled is a moat with a hazard rate. |
| A verification record that carries its own arguments | ≈ 0 | Not by policy, by construction. A result whose triple is bound to it can be re-run by anyone and returns the same answer; it cannot be taken by expenditure, because there is nothing to take. A rival may run more evaluations, faster, on more models — flow rents — and this changes nothing about which record reconstructs. It satisfies Definition 1 in the only way anything ever does: the transfer map is empty. |
Which returns the Plan’s own introduction to it. “Whoever has the largest AI ecosystem will set global AI standards.” CHECKED That sentence treats the standard as a prize of the race — won by scale, and therefore, on its own logic, losable to greater scale. WP-38 says a hub reached that way carries the hazard of the route that reached it. The standards layer is the one asset in the document with λ near zero, and the Plan proposes to arrive at it by velocity, which is the strategy for the case where λ is high.
WP-38 commits to a closed form, λ̄ = ((1−δ)/δ)·γ/(1−γ), and to a value of 0.0186 per period at δ = 0.985, γ = 0.55. That number is not carried across to this reading and no substitute is offered. It is calibrated to a financial position with a measurable discount factor and a defined private signal; geopolitical “patience” and “private information” are not those quantities, and the mapping of γ in particular is unresolved — a reference implementation is not a private signal.
Importing the figure would be OVER-GENERALISED in the sense WP-97 named one paper ago: a parameter fixed in the hypotheses, read in prose as a constant of nature. What transfers is Definition 1 and the question it licenses. The ordering in the table above is qualitative and is offered as such. OPEN
The practical reading is short. A race is winnable and must be re-won; the Plan is a good race document and this note takes no position on the racing. But the asset it names as the actual prize — the standard — is the one thing in it that could be held, and holding it is not a speed problem. It is the provenance problem of §§2–4, and whoever solves that first does not have to solve it again.
§7 asks which assets have a preemption hazard near zero. That question cannot be answered from one side. λ is the probability that some other player’s action transfers the hub, so it is a property of the other players’ action sets, not of the holder’s intentions. A contestability analysis with one player in it is not an analysis.
So: who does the Plan say is on the board? The document was searched in full — 28 pages, 69,187 characters of extracted text. MEASURED
| Term | occurrences | note |
|---|---|---|
| Europe / European / EU / Brussels | 0 | the jurisdiction with a comprehensive AI statute in force does not appear |
| “AI Act” | 0 | the three apparent hits are the string inside “AI Action Plan” and “Paris AI Action Summit” |
| ISO / IEC / JTC | 0 | the technical committee where AI standards are actually drafted is never named |
| Japan / Korea / India / United Kingdom | 0 | no ally is named anywhere in the document |
| China / Chinese | 6 | the only counterparty named at all |
| ally / allies / partner | 42 | the board is described in roles, not names |
| adversar… | 19 | likewise |
| NIST / CAISI | 34 | the instruments are named in detail; the field they act on is not |
The bodies that are named, once each, are the United Nations, the G7, the G20, the International Telecommunication Union and ICANN (p. 20) — diplomatic and internet-governance fora. The standards work the Plan asks NIST to lead happens in a technical committee the Plan does not mention.
It does not refute the ordering; it means the ordering was read off one player’s asset list. A hazard rate is estimated from what others can do, and this document supplies that information for exactly one other party. §7’s table is therefore a hypothesis about λ, not a measurement of it, and is corrected here to say so. OPEN
The Plan’s allies are not asked; they are aligned. Partners should “follow U.S. controls, and not backfill”, and where they do not, “America should use tools such as the Foreign Direct Product Rule and secondary tariffs to achieve greater international alignment” (p. 21). CHECKED The technology-diplomacy plan is to “induce key allies to adopt complementary AI protection systems”.
In WP-38’s terms that is V making the alliance part of its own cost structure rather than part of its position. It may well work on the assets where velocity is the right posture. But on a hub whose value is adoption — and a standard has no other value — every coerced adopter is an actor with a raised incentive to find an alternative reference. That is a mechanism that increases λ on the United States’ own strongest positional asset, which the Plan elsewhere identifies correctly: “leverage the U.S. position in international diplomatic and standard-setting bodies” (p. 20) concedes that the position already exists.
The middling λ assigned to compute above was reasoned the ordinary way: fabs and grid are years of carry, chips are transferable, Pillar III is loophole-plugging, so the hazard is real but slow. That reasoning takes the control’s target for granted. There is a second mechanism, and it does not involve anything leaking.
Export controls are aimed at the frontier — compute sufficient to train and run the most capable systems, with location verification on advanced chips, new controls on semiconductor sub-systems, and the Foreign Direct Product Rule behind both (pp. 21). CHECKED The control binds only where frontier capability is what the task requires. Where an already-diffused model is sufficient, the control is patrolling a margin that is not the constraint.
On a task with a cheap, sound checker, capability and verification substitute for each other. A weaker generator that proposes many candidates, filtered by a check that cannot be fooled, produces the same warranted output as a stronger generator, because the warrant comes from the check and not from the generator. Formal mathematics is the clean case: the kernel does the discriminating, and it does not care which model wrote the term.
On a task with no cheap checker — which is most of them, and includes the CBRNE and cyber risk assessment the Plan asks CAISI to lead on p. 22 — nothing substitutes. There is no filter to put in front of a weaker model, so capability is the whole of it and the frontier matters exactly as much as the Plan says.
Which sharpens §7 rather than overturning it. The Plan controls compute uniformly and does not distinguish the two regimes, so on the asset this note is about — the one with the lowest λ — the control does not reach, and not because it leaks. It does not reach because the work does not need what is being controlled. A verification hub is serviced by a checker and a sufficient generator, and “sufficient” on that task was reached some time ago.
That is also the same sentence as §8’s labour subsection and WP-99 §6, arriving a third time. The property that makes the record self-warranting is what decouples it from the thing everyone is racing for: from the credential, from the language, and here from the frontier.
The substitution rate is unmeasured. How much weaker a generator can be, on a task with a sound checker, before the total cost of getting a warranted result rises — that is the number that decides whether an export control binds, and this note does not have it and has not seen it. “Older models do fine” is a claim about a saturated benchmark until someone states the task and measures the frontier.
And there is a term running the other way. A weaker generator may need many more attempts, so substitution buys capability with inference, and the trade favours the frontier wherever compute is cheap and favours the checker wherever the check is cheap. On a kernel check — seconds, deterministic, no human in the loop — the check is very cheap indeed, which is why the clean case is the clean case. It does not generalise on its own. OPEN
The previous subsection asks how good a model has to be. That is the wrong question one level up, and the right one is older than this subject. It is possible to recycle water indefinitely and grow potatoes forever. Possibility settles nothing. The question is whether the scarce input is worth more in the potato field or somewhere that a marginal unit changes an outcome — recovery after neurosurgery, to take the case where the divergence is starkest.
This is WP-41 of this volume, in another domain. Planetary Triage opens on the observation that $110 trillion of global product cannot simultaneously relocate 2.3 billion displaced people, fund transformative decarbonisation and deploy atmospheric intervention — so the choice is forced, and a strategy that does not choose has chosen by default. The same shape applies here, and the analogy is exact for an unobvious reason: the contested input is fungible. Capability is not transferable between tasks, but compute, energy, capital and trained people are, and those are precisely what the Plan is about. Allocation is a real question rather than a category error.
On a saturated task — one where the requirement, or the checker, was already met — the marginal return on additional capability is near zero. Every unit spent there buys nothing, and the units are not free: they are drawn from a common pool of compute, grid and people.
On an unsaturated task the marginal return is high. The allocative question is whether the strategy directs the marginal unit toward the second kind, and it is answerable from the document.
The Plan is not silent on direction. It names healthcare, energy and agriculture, and launches domain-specific efforts through NIST (p. 5); it asks for automated cloud-enabled labs across engineering, materials science, chemistry, biology and neuroscience (p. 8). CHECKED But read for instrument rather than for intent, those are convening, standards and data-access mechanisms. The instruments that actually move compute are two: a financial market for compute and the NAIRR pilot (pp. 4–5), which allocate across actors — American startups and academics rather than incumbents — and export controls, which allocate by denial across borders. Neither allocates across tasks. Within the actors it favours, the market decides, and a market allocates by willingness to pay.
Which is where the potato and the neurosurgery separate. Willingness to pay tracks appropriable return, not the return itself, and those diverge most sharply in exactly the places the second example points at: recovery, rehabilitation, diagnosis at the margins, the long tail of conditions with no volume behind them. That is not a criticism unique to this Plan — it is the standard result about markets and public goods — but a strategy whose only allocation mechanism is a market has inherited the result along with the mechanism, and the document does not say so.
You cannot triage what you cannot see. Saturation is only visible after the fact: the tasks where capability stopped mattering are known now, and the ones where the next increment unlocks something are not. General capability is general precisely because it transfers to problems nobody listed in advance, so racing the frontier is a hedge against the allocative question being unanswerable ex ante. WP-41 could triage because its constraint set was measurable — hydrology, temperature, sea level. Here it is not, or not yet.
Both readings are consistent with the document, and this note does not decide between them. What it observes is narrower: the Plan makes the hedge without stating it as one, and a hedge that is not stated cannot be reviewed, priced, or revisited when the saturation map fills in. OPEN
And the observation that would fill it in is the same one the previous subsection needs: a map of tasks by whether a cheap sound checker exists, and by the marginal return on capability given one. That is a measurement, it is nobody’s in particular, and every allocative argument here — the Plan’s and this note’s — is waiting on it.
A standards hub is held by whoever is conformed to, and conformity is a population fact. The Plan’s own leader-list is supply-side: “America currently is the global leader on data center construction, computing hardware performance, and models” (p. 20). Standards are not on that list — the one asset in §7 with low λ is the one the Plan does not claim to lead.
Set against that, the largest organised bloc outside the alliance the Plan describes is BRICS, which by 2026 reporting is variously given at roughly 40–44% of world GDP and 45–56% of world population depending on whether partner states are counted and whether GDP is nominal or PPP. MEASURED Those ranges are wide because the grouping’s membership moved recently and sources differ; the ranges are given rather than a point estimate for that reason.
If a hub is held by adoption and cannot be bought, then the decisive variable is the size of the population that conforms. On that reading a strategy that names no ally, coerces the ones it has, and treats the majority of the world’s people as neither ally nor adversary is not underspecified about the board — it is contesting a hub while conceding the demand side. The remedy is coalition rather than velocity: join the adopters instead of racing the rival.
And the case against it. BRICS is not a standards-setting body, has no AI conformity machinery, and does not act as a unit on technical questions; population share is not adoption. A bloc can be large and still not be a hub. The argument above establishes that the demand side is uncounted, not that any particular coalition holds it. OPEN
The Plan’s human-capital sections are about displacement and about building the physical plant: “Empower American Workers in the Age of AI” (p. 6) is retraining for people affected by adoption, and “Train a Skilled Workforce for AI Infrastructure” (p. 17) is the trades that build data centres. Across the whole document: workforce 22, apprentic… 5, educat… 15, university 0, K-12 0, visa 0, immigration 0. MEASURED A strategy for a talent-intensive contest contains no education pipeline and no immigration instrument.
Set that against the arithmetic of the two largest populations. Suppose India and China each brought 30% of their people to a computer-science education — the hypothetical is not a forecast, and its implausibility is discussed below.
To match on headcount, that bloc would need eighty-three percent of every person in it — infants, retirees, everyone — holding a computer-science education. There is no policy that reaches that number, which is the sense in which the comparison does not close. MEASURED Inputs and the division are in wp98-verify.py so a reader who disputes a population can move it and see what happens.
Thirty percent of a total population holding a CS education exceeds total tertiary attainment in every country that exists, so read as a schooling target the premise is a bound rather than a plan, and neither state has announced anything of the kind. Nothing here forecasts that they will. But that reading measures credentials, and the subsection below argues credentials are the wrong proxy — which is a defect in this paragraph, corrected there rather than deleted here.
Headcount is also not frontier capability. Frontier research is done by small teams, and the binding constraint on it has historically been compute and capital, not people. On the frontier-model asset — the high-λ one in §7 — this arithmetic says little. OPEN
Where it does bite is the asset this note is about. Conformity assessment, evaluation, audit and proof engineering are labour-intensive, distributable, and bounded below by how many people can do the work — unlike training a frontier model, which is bounded by capital. A verification hub is serviced by headcount in a way a model lead is not. So the ranking in §7 has a mirror: the asset with the lowest preemption hazard is also the one whose input is the least purchasable, because you cannot buy a population of assessors on the timescale of a procurement.
That is the sharpest form of the demand-side point, and it does not require any coalition to act as a unit. It requires only that the work be doable by many people in many places, which is what distinguishes a standard from a model.
The bound above says a CS education for 30% of a population is out of reach because no education system has ever produced attainment at that rate. That argument holds only if credential is the measure of who can do the work. It is a parameter fixed inside the analysis and read as a constant of nature, which is the OVER-GENERALISED shape WP-97 named — here, in this paper, one section after being warned about it.
What makes it a defect specifically for this hub is §5. A gate report is evidence because it does not depend on who ran it. That is the whole design: the number cannot be typed, so it does not matter whose fingers were near the keyboard, and a third party re-runs it rather than trusting a reputation. The property that makes a verification hub non-contestable is the same property that makes its labour supply credential-independent. A machine-checked result is the rare kind of professional output that carries its own warrant, so the person producing it needs the discipline and the tooling, not the diploma that historically certified both.
If that holds, the binding input is not schooling. It is devices and access — hardware in hand, and a model to steer with. Which is a constraint that moves in quarters rather than in generations, and moves fastest exactly where population is largest. The Plan is not silent on compute access; it proposes financial markets for compute and the NAIRR pilot (pp. 4–5). Both are scoped to American startups and academics. There is no analogous instrument in the document for anyone else, and no reason the same mechanism could not be built by anyone who wants the assessors.
This paper is one. Its corpus figures come from a Lean 4 development with 148 kernel-audited declarations and eleven proofs resting on no axiom at all, produced by an author with no computer science degree, steering a model. The four defects in §6 were found by instruments a non-specialist specified and a non-specialist runs.
And its limits, which are severe. n = 1. The author holds three baccalaureate degrees and two decades of professional analytical work, so the steering landed on someone already trained to interrogate an answer — which is not the same as trained in the field. The record also shows the correction traffic running the other way: the machine produced all four defects in §6, and the corrections that found them — a Grönwall bound misdescribed as a chosen threshold, a grid mistaken for a fact about six and ten, a board with one player on it — each came from the human. A steered non-specialist is not an unsupervised one, and this note makes no claim about the second. OPEN
So the labour claim narrows and, narrowed, gets stronger. Not AI makes everyone an engineer. Rather: on a hub whose outputs are machine-checkable, the credential requirement falls to near zero, because the artefact carries the argument instead of the author carrying a reputation. That is the same sentence as §2, applied to who may hold the position rather than to what the position is. It is why the hub cannot be bought, and it is why it cannot be fenced.
Three observations, none of which this note has, and each of which is available to someone who looks:
Until those are on the table, §7 stands as a hypothesis with the sign of its own uncertainty recorded. The instrument in §5 is unaffected either way: a record that carries its own arguments is worth building whoever ends up holding the hub, and is the one item here that does not depend on knowing.
Not a proposal. A statement of the minimum an interoperable evaluation record would have to contain for a second party to know what the number is a number of.
It does not claim that formal verification of Lean theorems and evaluation of frontier models are the same activity. They are not, and the second is enormously harder: a theorem has a statement and a model has a distribution of behaviours. The transferable part is narrower and it is the provenance discipline, not the mathematics — how a result is bound to the conditions that produced it, and what a checker must enumerate to be sound.
It does not claim the instruments here are policy-grade. They are one person’s tools, running on one corpus, and §6 is a list of what they found wrong with that corpus and not a record of adoption anywhere.
It takes no position on the Plan’s other pillars, on its energy and permitting provisions, or on its direction to revise the NIST AI Risk Management Framework (p. 4) — noted here only because that framework is the artefact the evaluations ask on p. 10 would build on, so its scope is a fact about where this work could land. CHECKED
And it is not a bid. The corpus has one federal proposal in flight and this is not a second one.
§12 argues from the diffusion record of earlier general-purpose technologies. There is one respect in which the present case is not like them, and it is the speed.
| Technology | to 70% US households |
|---|---|
| Telephone | ~60 yr |
| Electricity (from 1900) | ~40 yr |
| Technologies arriving by 2000, trend | ~17 yr |
| and, on the different measure noted below | |
| ChatGPT to 100 million users | ~2 mo |
| … to 10% weekly US users | < 2 yr |
| … 10% → 30% | ~6 mo |
RECORDED The comparison is not clean and the source that assembles it says so: the historical series measure household penetration and the recent ones measure individual users. Those are different units, which is the defect this whole paper is about, appearing in the one table that would most like to be a clean one. The ordering survives the caveat; the ratios should not be quoted as if they did.
Faster diffusion is not a counterexample to §12. λ is a hazard rate, so a shorter diffusion time is a higher λ — the half-life of any lead in the technology itself is shorter than it has ever been. Speed is indifference running faster, and it sharpens the ordering in §7 rather than disturbing it. It also shortens the window WP-94 identified: a standardisation window is open while the layer that would need the standard does not exist, and a compressed adoption curve closes that window on a compressed schedule.
§12 lists, among the positional advantages that survive when a technology diffuses, a unit of account. That is not one example among several. It is the largest instance of the object this paper has been describing, because a unit of account is the standard that every other standard is priced against.
The proposition put to this note was that the decisive asset would be a “true fiat money — a trusted and apolitical store and transfer of wealth”. Fiat is money by decree, so apolitical fiat is a contradiction in the term rather than in the idea: what is described is a non-sovereign settlement asset. The distinction is recorded because precision costs nothing here, and then set aside — twice in this series a correct structure has arrived in borrowed vocabulary and been argued with instead of followed (WP-97 §8), and the structure is the point.
The structure is §2, at the scale of money. A sovereign currency’s value is a claim whose argument does not travel with the token. The token names itself; the things that make it worth anything — issuer solvency, legal enforceability, the stability of a political order, the discipline of a central bank — are recorded elsewhere, unlinked to any particular unit in any particular hand. That is the triple problem exactly: the artifact is carried and the arguments are not.
And the λ follows from it. An issuer-backed unit has λ > 0 for a reason that is not speculative but historical: position transfers when the issuer’s standing does, and it has, more than once. A unit whose warrant was self-carrying — verifiable by the holder, not vouched for by an institution — would be the same object as a record that carries its own arguments, and would have λ ≈ 0 for the same structural reason. That is a real observation about hub properties and it is the strongest form of this paper’s thesis.
No such asset exists. Many have been proposed and offered as this; none has become a unit of account at any scale, and the paper takes no position on any of them, endorses none, and is not investment advice of any kind. The author is not an advisor and this note is not a recommendation to hold, transfer or price anything.
The premise may be self-defeating. There is a serious argument that a unit of account requires the political apparatus the proposition removes: a lender of last resort, legal enforceability of contracts denominated in it, and an elastic supply without which the unit is a store of value that is bad at being a unit. On that reading “apolitical” is not an improvement on money, it is incompatible with one of money’s three jobs.
Low λ is not fast displacement. Monetary network effects are the strongest that exist. A hub can be theoretically uncontestable and practically immovable at the same time, and the incumbent’s inertia is a fact about adoption rather than about hazard. Nothing here predicts a transition, a timescale, or a winner. OPEN
What survives all three objections is narrow and is the only part claimed: the property that would make a settlement asset non-contestable is the same property that makes a verification record non-contestable — the warrant travels with the thing instead of living in an institution that has to be trusted separately. Whether anything can hold that property at the scale of money is outside this note. That it is the same property is not.
A settlement asset is a lingua franca in the narrow sense: a medium in which parties who share nothing else can transact. But the word means a language, and a language is not restricted to the first thing it is used for. Latin spread as administration and stayed as scholarship; English travelled on trade and became the medium of science. Systems that begin transactional do not remain transactional, because a medium that reaches everyone is the cheapest medium for everything.
Which locates money correctly as the thinnest possible instance of this paper’s object. A monetary unit transmits one thing: a quantity. It carries no argument, and it is not supposed to — that is what makes it fungible. A verification record transmits the quantity and the reason. It is the same kind of object, strictly more expressive, and it is what the medium turns into if the pressure runs the way media have always run.
The position on such a medium is held by whoever has custody and transmission — not by whoever generated the content. This is the ordinary history of media: the protocol outlives the power that established it, the notation outlives the civilisation, the journal outlives the academy. The custodian of transmission is the durable role.
And the ordering claim, which is the sharp one: knowledge precedes power where the medium is concerned — custody of what can be transmitted is upstream of what can be enforced, not downstream of it.
The case against that ordering is strong and has most of the history in it. Transmission media have usually been established by power and inherited afterwards: Roman roads before Latin’s reach, sea power before English’s, postwar institutional design before the standards that ride on it. On that reading power precedes knowledge and the durability of the medium is just the long tail of a settlement someone won. Nothing in the historical record decides this cleanly, and this note does not decide it either.
What it can say is narrower and is new rather than historical. Every transmission medium so far has required a trusted custodian — and a custodian is credible in proportion to the power standing behind it, which is exactly why the ordering has run power → knowledge. A medium whose records verify themselves does not need one. Custody stops being a position of trust and becomes a position of maintenance. That is the first structural circumstance in which the ordering can run the other way — not evidence that it will, but the removal of the thing that prevented it. OPEN
The proposition was put that the lingua franca need not be English. The stronger form is that on this layer it need not be a natural language at all, and cannot usefully be one.
A kernel checks structure. A proof term means the same thing whatever the surrounding prose is written in; the report in §5 that says a declaration rests on no axiom says it identically to a reader in any language, because the sentence is not what is being checked. That is the same indifference as §12 and §8, in a third place: a gate that does not check a passport does not check a language either.
Which is a practical asymmetry and not a sentiment. A lingua franca made of prose has to be learned, and learning it is a cost paid by everyone who is not a native speaker — a permanent transfer to whoever already speaks it. A lingua franca made of machine-checkable records has no native speakers. The cost of entry is the tooling and the discipline, which are purchasable in quarters, rather than the language, which is not.
The Plan closes its introduction: “The AI race is America’s to win, and this Action Plan is our roadmap to victory.” (p. 2) CHECKED Every argument in this note has accepted that framing and asked only which assets survive inside it. This section says what the framing itself assumes.
A race presumes the prize can be held against the other runners. General-purpose technologies do not behave that way, and the record on this is long and one-directional. Steam, electricity, the transistor, the shipping container, the protocol suite underneath the internet — each was developed somewhere particular, and in every case the developer’s lead in the technology decayed toward the world average, while whatever positional advantage survived was institutional: a standard, a protocol, a legal regime, a unit of account. The thing itself has no allegiance. It works identically for whoever holds it, which is precisely what makes it general-purpose.
That observation is not an objection to the Plan — it is the argument of §7 arriving from outside the model, and it points the same way. The high-λ assets are high-λ because the technology is indifferent. Weights diffuse for the same reason electricity did.
A hub that cannot be taken by expenditure also cannot be fenced by it. If a record carries its own arguments, then anyone who runs it gets the same answer — which is what makes the position unbuyable, and equally what makes it unownable. You cannot lose it, and you cannot hold it to the exclusion of anyone else.
So win is the wrong verb for this one asset. The right verb is occupy, and occupation is not exclusion. The standard is not a prize taken from a rival; it is a commons someone has to actually build, and the builder holds the position by having built it. That is a strategy available to a party of any size, which is the only reason this note exists.
The instrument shows the same indifference at the smallest scale, and it is worth noticing because it is the same property. A gate that checks a proof does not check a passport, an institution, or an intention. In §6 it refused its own author four times in a week, and in §8 it refused this paper’s own arithmetic and printed the correct figure. That is not a tool being rigorous. It is a tool having no side to take — the same fact as diffusion, observed at the scale of one file.
Which leaves the framing where it should be left, descriptively. A transformation that takes no sides does not thereby take yours. It confers no standing on the party that assumes it is theirs to win, and withholds none from the party that does not. What it responds to is the same in both cases: whether the work was done, and whether the record shows it.