The Legitimacy Gate

I. The legitimacy gate

You can perfectly compute the failure mode of a government policy, price it in a prediction market, and publish it for free. The policy will still happen.

A prediction market says the policy will fail. An AI audit finds the contradiction. A think tank publishes the mechanism. A spreadsheet proves the fiscal path is impossible. Nothing happens.

This is the standard failure mode for the modern repertoire of governance fixes. The signal is correct. The signal is public. The signal does not bind, does not steer, does not get answered, and within a year it is forgotten. The default diagnosis is either public choice (“politicians are corrupt”) or epistemic inadequacy (“voters are irrational”). Both are sometimes true, and public choice explains a good deal about why the missing step stays missing: no incumbent gains from a procedure that binds them. Neither says what the step is.

The missing step is what I’ll call the legitimacy gate: the conditions an artifact’s institutional attachment must meet before the artifact can bear on a public decision. Five questions put those conditions in diagnostic form, and not one of them is about the artifact’s content: a market price, an audit finding and a full consequence model all face the same five.

• Who authorized this model?
• Who can contest it?
• Which institution owns the response?
• What happens if it is ignored?
• Which legitimate decision-maker must take it into account?

An artifact passes the gate when all five have answers: a named authority stands behind the model, a named party can contest it, a named institution owns the response, something specific happens if it is ignored, and a named decision-maker must take it into account. In operation that means a visible procedure — endorse the model, contest it, override it, delay the decision, trigger an audit, or schedule a public review comparing the prediction to what actually happens. Every one of the five is a duty someone holds, not an incentive someone feels. Answer four of the five and the artifact is still advice.

Earth’s state machinery is legitimacy-first. The apparatus that answers who may decide — constitutions, mandates, appointment rules, courts, elections, the doctrines fixing when an official may act — is the legitimacy stack, and it got centuries of institutional iteration. What will the decision do did not. The asymmetry is the spine of Legitimacy Came Before Cognition; this essay is its applied sibling. A signal can be true, useful, public, and still have nowhere legitimate to attach. When that happens, it becomes advice, journalism, activism, or a spreadsheet that someone politely fails to read.

II. Specimen: every actor was acting within role

The 2022 Statistics Finland reclassification of state-subsidized housing loans (developed in detail in Legitimacy Came Before Cognition) is the canonical case: every actor acted within role, the classification was procedurally legitimate, no institution was required to model the consequence. Finland’s headline general-government debt ratio rose by about six percentage points overnight. The classified stock has since grown toward €20 billion. Whether the reclassification is also why the country’s primary countercyclical housing instrument is now politically harder to use, just as construction enters a deep recession, is not something the accounting record settles — and it was nobody’s job to find out.

Every actor was acting within role. That was the failure.

A prediction market could have helped only if someone had formulated the right question and attached it to the decision before the classification locked in. That formulation step — selecting which question becomes institutionally live, who must answer it, and which decision-maker must take the answer into account — is what passing the legitimacy gate would have required. No Finnish institution was required to perform it. Where else the shape should appear, if the diagnosis is right: procurement decisions, regulatory classifications, fiscal-rule interactions, AI deployment thresholds — anywhere a legitimate procedure produces a consequence nobody was required to model, and the cost is absorbed downstream.

III. Prediction markets are inputs, not the cognition layer

A prediction market answers priced questions: will X happen? It does not select which questions become institutionally live. It does not specify the mechanism a policy claims will produce X. It does not name the hidden variable that absorbs the cost the policy displaces. It does not assign ownership of the response. It does not impose a duty to act when the price moves. A market is signal infrastructure.

A cognition layer, in the sense I mean, has two halves: it produces consequence models, and a named decision-maker must answer them. For a proposed decision, the modeling half outputs:

1. The mechanism claim — the causal path that is supposed to produce the stated outcome.
2. The carrying assumptions — what must be true for the mechanism to work.
3. Forecasts — what we should expect to observe and when, including market and expert signals where available.
4. An absorption audit — which variables nobody has to answer for, and may therefore carry the cost.
5. An owner map — which institution must answer for which downstream consequence.
6. A falsification trigger — what observation would show the mechanism failed.
7. A review date — when the prediction is compared to outcome in public.

Prediction markets are useful inputs to (3). They do not by themselves produce (1), (2), (4), (5), (6), or (7). The attachment half is the legitimacy gate doing its work: a specific decision-maker must either publicly endorse a falsifiable mechanism model — with its assumptions, its absorption audit, and its review date — or have the decision delayed, contested, or marked as unmodeled in a place that subsequent voters can see. “Attach to a decision” means that level of teeth, not “publish a PDF appendix.”

A note on the relation to Constructive Diagnosis: that essay’s six-field repair specification is the architectural standard for designing the institution that holds the missing primitive. The seven outputs above are what the institution, once built, produces for a given decision on a given day. Specification and output are different artifacts.

The civic-tech version of this problem. The broader belief that public data, transparency dashboards, prediction markets, AI audits, and forecasting tournaments will collectively translate into better governance runs into the same gate.

Hanson’s futarchy — vote on values, bet on beliefs, where representatives define a welfare metric and markets estimate which policies raise it — is the cleanest existing attempt at exactly the move this section names. Constitutionalizing the market’s role inside a values metric set by legitimate principals answers two of the gate’s five questions by construction: a named principal authorized the model, and the decision-maker who must take it into account is fixed in advance, because the verdict attaches to policy automatically. Contestation and override are left open, and so is who sets the agenda the market prices. On the attachment axis, futarchy is the cleanest existing proposal.

Run it against the seven outputs above and a different picture appears. The market produces (3), the forecast. It does not produce (1) the mechanism claim — traders are not required to publish their causal models, and the price aggregates private heterogeneous beliefs into a probability without naming any of them.

It does not produce (2) the carrying assumptions, which stay private; or (4) the absorption audit, because no one writes a market on a variable nobody has named. (5), (6), and (7) are partial: there is constitutional ownership at the meta-level and price-vs-realization comparison at settlement, but no per-decision owner map and no mechanism-level falsification, because there is no published mechanism to falsify. Futarchy binds without modeling. It is a binding forecast layer, not a cognition layer. Constitutionalizing the market makes its verdict authoritative; it does not make the market produce the missing outputs.

A bounded oracle, as I’ll use the term, is an institution that produces decision-relevant cognition inside a scope-limited mandate set by a legitimate principal — it answers a defined class of questions for a defined client, and stays out of the rest. The bounded oracles that did emerge on Earth — the Congressional Budget Office, central banks, fiscal councils — took that shape and survived. They did not try to make cognition the general principal; they took a bounded mandate from one.

Run them against the same seven outputs and the picture is mixed. A cost estimate states a mechanism, its assumptions, and a projection for a bounded class of questions — more than a market price supplies. It carries no absorption audit, no per-decision owner map, and no mechanism-level falsification trigger. What these bodies solved was attachment: a named client must receive the number and legislate in its presence. Nothing in their mandate asks for the rest.

Futarchy’s open problems — who defines the welfare measure, who controls the policy agenda, which forecasts are admissible, how manipulation is audited, how overrides work, why losing factions treat the result as legitimate — are not engineering edge cases. Every one of them is a question about who holds the principal’s seat, which is the question the legitimacy stack exists to settle and the one thing futarchy proposes to hand to a market.

Binding futarchy therefore appears, if anywhere, where no incumbent faction had to consent to it — a new charter, a small organization, a jurisdiction written from scratch — and not in an established state. Not because it is incoherent. Because no incumbent faction gains from cognition selecting policy over them.

IV. AI eats the technical bottleneck, not the legitimacy bottleneck

AI and open-source tooling dramatically lower the cost of producing cognition artifacts. They do not solve the procedural bottleneck. Producing a true, computable diagnosis outside the state does not make it decision-relevant. AI does not decide which artifacts are legitimate, which must be answered, which decision-maker must change course, or which institution must own the response. It changes the political question from “can the state compute this?” to “which legitimate actors must answer when the computation is public?”

This is also where the most leverage sits in the next decade. The diagnostic substrate built outside the state is exactly the right move. Tools that compute statute-graph diagnostics — stale citations, unexercised mandatory delegations, orphaned decrees, structural drift between in-force and as-enacted text — produce findings from public data without needing state permission. Whether such findings change anything depends on whether some legitimate actor is obligated to respond.

Formal inadequacy is now a computable object: those defects fall out of public data by machine. Consequence is not. Seeing the 2022 reclassification coming needed a model of what one classification decision would do downstream, which no statute-graph tool produces. And repair is not computable at all.

V. Design constraints: Hayek, Scott, and the NEPA trap

Cognition layers can fail in two opposite directions. Hayek’s knowledge problem and Scott’s high-modernism critique name the first: the layer mistakes its model for the world, and centralized confidence destroys local information. The second is that the layer survives but becomes ceremonial.

The US National Environmental Policy Act (NEPA) and its environmental impact statement (EIS) regime is the cautionary case for the second. NEPA created a procedural requirement for federal agencies to review significant environmental consequences and inform the public before major decisions — recognizably a consequence-modeling requirement. In practice, final EISs commonly run to several hundred pages and take years. The Council on Environmental Quality’s 2013–2018 sample found a 661-page average and 447-page median final EIS, well above the original 150–300-page expectation; recent CEQ data give a median of around 2.2 years from notice of intent to final EIS.

The result is documents that function partly as litigation defense and compliance documentation rather than as agile model-update infrastructure. Power’s Audit Society generalizes the same failure mode: when verification becomes mandated, capture answers with rituals. This is why filing a model in a public register does not by itself pass the gate: a register entry answers nobody, and the gate’s fourth question — what happens if it is ignored — has no answer for a document that only has to exist. Both Hayek/Scott and NEPA accept the premise and attack the execution. They are design constraints, not refutations.

A cognition layer that survives both must:

• Publish uncertainty. Publish assumptions, uncertainty, causal paths, excluded variables, and expected absorption points.
• Import local knowledge. Import local knowledge without pretending local knowledge has been fully centralized.
• Compare prediction to outcome. Compare prediction to outcome in public after the decision.
• Separate model from decision. Carry adversarial review and a right of reply.
• Expose model error. Have the contestation route mandated, not just the production route — so that compliance-theatre fictions can be challenged by counter-models with formal standing.

Designing around the NEPA trap also requires hard procedural bounds rather than open-ended compliance: strict page and time limits, automated rejection of boilerplate language, mandatory machine-readable structure, and non-trivial adversarial-review funding so that opposing-model production is not orders of magnitude underresourced relative to the official model.

VI. Build the diagnostic substrate first

The mistake is to start with “mandate this in legislation.” That asks the equilibrium to have already solved itself: asking the inadequate system to bind itself by passing a Mandatory Cognition Act is asking for the legitimacy gate to already be open. A mandated contestation route is still the end state — a cognition layer without one gets captured. It is starting there that fails. Two paths avoid the trap — build signals that need no permission, and build procedures that ride existing legitimate vehicles.

The honest framing for the unofficial artifacts is that they are catalysts, not the cognition layer itself. A one-shot think-tank PDF gets ignored because it is a one-shot signal with no compounding mechanism and no cheap path for any rival institution to use. An automated, versioned, public diagnostic substrate is different in three specific ways:

• It compounds across years and electoral cycles, accumulating receipts faster than any individual decision-maker can outrun.
• It indexes durably against named decisions and named decision-makers, so the cost of a rival rediscovering the inadequacy collapses to a search query.
• It lowers the activation energy for legitimate actors with grievances — opposition staffers, courts, regulators, journalists, fiscal councils — to wield the diagnosis, the way opposition parties, courts, and journalists have long wielded independent think-tank and academic analysis against incumbents, but at automated cost, indexed durability, and structural specificity that previous artifacts could not match.

The substrate doesn’t bypass the legitimacy gate. It catalyzes the actors who already have keys to it.

This is also where statute-graph diagnostics diverge from prediction markets. A prediction market produces a probability: “this policy will fail with X%.” As ordinarily produced — outside a constitution that binds someone to it — that number has no procedural attachment, no specific contradiction, no citeable structure.

A statute-graph diagnostic produces something different — “section A requires the government to issue regulation B; regulation B has not been issued; the legal entitlement under section A is therefore unimplementable in cases C and D” — which is an artifact opposition staffers can drop into a parliamentary motion or that courts can incorporate by reference. Probability and legal weapon are not the same kind of object. Markets price questions; substrates produce structured, citeable findings that already speak the legitimacy stack’s native language.

Outside-the-state moves (no permission needed).

• Open-source legal-state replay. Tools that compute stale citations, orphaned decrees, unexercised mandatory delegations, and statute-graph drift per jurisdiction, published before each electoral cycle. The inputs are public records, so producing the finding needs nobody’s authorization. Attaching it to a decision still does.
• Counter-model registries: public, versioned consequence models for major bills, produced unofficially by think tanks, civic groups, or AI-augmented analysts, attached to the bill’s name in a way that journalists, opposition staffers, and courts can find later.
• Forecast-to-outcome receipts: anyone can publish the receipts when an official forecast is wrong by more than its stated uncertainty band, with the receipts attaching durably to the decision-maker’s name in indexable public records.

Pre-existing-vehicle moves (require legitimate principals, not new agencies).

• Generalize budget-office scoring beyond fiscal cost to mechanism claims, inside the existing fiscal-council mandate. The legitimacy substrate already exists.
• Attach mechanism briefs to ministerial impact assessments where existing law already requires impact assessment. Use the existing legal hook; expand the content.
• Use bounded oracles’ existing remit to publish consequence models for specific upstream inputs to their domain (e.g., a central bank’s macroprudential mandate may already justify publishing consequence models for housing-finance classifications that affect debt ratios and construction cycles).

These artifacts do not yet bind decisions. What they change is the next argument: the question shifts from “could anyone have known?” to “why was the known model ignored?” That is the lever by which inadequate equilibria start to move.

VII. A falsifiable prediction: AI governance bodies

Take the AI safety institutes, AI Act enforcement bodies, and national AI governance organs operating at the end of 2026. Score each on three measures fixed now rather than after the fact:

• Standing — is the body named in primary legislation, with some other actor placed under a duty to respond to its output?
• Uptake — are its findings cited in another institution’s formal decisions: enforcement actions, court filings, budget documents, licensing conditions? Press coverage and the body’s own reports do not count.
• Technical quality — does its evaluation work survive external review or reproduction by parties that do not fund it?

By the end of 2031, count which of the cohort still hold a budget line and a mandate under their own name. The prediction is that standing and uptake sort the survivors and technical quality does not: a body scoring low on the first two and high on the third gets absorbed, renamed, or cut back to advisory status. The falsifier is a body with no statutory duty owed to it and no citations in another institution’s formal decisions that is still standing in 2031 with its mandate intact. Uptake is scored in 2026 and off a different record, so a survivor cannot be recoded afterwards as having been useful all along.

VIII. Close

An adequate civilization is not one where every citizen personally outthinks the state. It is one where the state’s decisions are attached to public, contestable consequence models before damage occurs.

Prediction markets are useful. AI is useful. Expert forecasts are useful. Open-source diagnostic substrate is useful. None of these is the cognition layer. A cognition layer is an institution that produces contestable consequence models which a named decision-maker must answer. Everything on that list produces. None of it is answered.

Signals are getting cheaper. The attachment problem is the work.

Related:

• Legitimacy Came Before Cognition — the historical asymmetry between legitimacy and cognition stacks; this essay’s parent
• Constructive Diagnosis — the methodological standard whose six-field repair spec this essay’s seven-output cognition layer instantiates
• Powerless Intelligence — the AI-era downstream variant: evaluators exist, response duty does not
• Bad Equilibria Are Not One Thing — the failure-mode taxonomy the cognition layer would diagnose
• The Fourth Branch — the institutional seat of the diagnostic operator
• Mechanism Realism — the foundational ontology the corpus runs on
• What Bureaucracy Is — the positive synthesis: bureaucracy as the runtime of traceable discretion; the legitimacy gate supplies the authorize function