This revision explains the customer value of the record: interpreting market-implied impact with market conditions, evidence and independent reviews. The historical observations and forecast register retain their original dates and claims. In particular, the September argument that a long outcome record could answer whether markets chose well needs a qualification: observations alone cannot identify the unchosen counterfactual. Decision quality and causal impact require an appropriate comparison and research design.
Key takeaways
- Capital is building intelligence at industrial scale. The IEA puts 2025 capital spending by five large technology companies above $400 billion, with a further 75% increase expected in 2026. As intelligence becomes cheap, the scarce resource becomes the decision: which plan to fund, who answers for it, and whether it was carried out.
- Intelligence multiplies options, not decisions. An organization with capable models can generate thousands of credible proposals and still fund only a few. Committees built for a world of scarce analysis cannot review them all.
- Decision markets turn disagreement into action. A decision market prices the same outcome under competing actions and lets the comparison choose. MetaDAO runs this in production on Solana today, where the price is the decision.
- An economy of intelligent agents will need markets to govern it. Once many independently controlled agents allocate real resources, central review stops scaling. Prices coordinate what no reviewer can see, and a market pays whoever corrects it.
- Prices need interpretation. Eidos explains the conditional comparison alongside proposal terms, participation, depth and separate review findings where evidence exists. The sourced record makes that interpretation inspectable. Eidos does not trade, propose, decide, hold funds or run a market.
- The evidence is early and we say so. No controlled, real-money comparison yet shows that a market chooses better than a committee. We set out what would prove us wrong, and on 29 September and 2 October 2026 we registered five dated forecasts against the thesis.
Capital is building intelligence
Three uses of capital are growing at the same time: building intelligence, deploying it in businesses that earn a return, and coordinating the decisions those businesses and their agents make. They form a chain, and the whole chain has to pay: sustained investment at the front depends on useful returns at the end.
The first link is already concrete. The IEA puts 2025 capital spending by five large technology companies above $400 billion, with a 75% increase expected in 2026, although it does not say how much of that spending is AI. It estimates that data centres used 485 TWh of electricity in 2025 and projects 950 TWh by 2030, close to a doubling in five years for a single physical input.
Financing is part of the constraint. Data-centre projects have outgrown what company balance sheets can fund, so their pace now depends on capital markets and on investors continuing to expect AI to pay. A useful project still waits if it cannot borrow on workable terms.
| Use of capital, 2026–2036 | Where it goes | What sustains it | What interrupts it |
|---|---|---|---|
| Build capacity | Compute, power, data centres, cooling and networks | Demand that keeps the capacity used and covers its financing | Overbuilding, power delays, expensive funding, weaker returns |
| Deploy capability | Businesses using AI for products, research, software and operations, and the data and payment rails they need | Measurable productivity, customer spending, retained margins | Deployment costs, unreliable performance, competition eroding the gains |
| Coordinate activity | Allocation mechanisms, evaluation, execution controls and evidence for organizations and agents | Enough consequential decisions to justify dedicated infrastructure | Central systems that solve it well enough, poor metrics, markets too thin to inform |
The table describes uses of capital rather than measured flows between them, because no such ledger exists. New financing can fund new activity, but buying an existing asset mostly pays its seller, bank lending creates deposits, and market capitalization can change with no matching inflow of cash.
The third row, coordinating activity, is the subject of this report.
Intelligence multiplies options, not decisions
AI enters the capital story as something expensive to build. It then becomes a tool that lowers the cost of software, research and operational work. As systems plan and carry out longer tasks, it becomes a participant: agents acting within budgets and mandates.
Each step adds an allocation problem. An organization with capable models can generate thousands of credible proposals and still have the capital, power, time and attention for only a few. More intelligence widens the set of options without removing the need to choose.
A capable model can recommend an investment, but the recommendation leaves several questions open:
- whose objective the model is optimizing
- what information it does not have
- how someone else can challenge its forecast
- who bears the downside
- whether the chosen plan is actually carried out
A more persuasive answer does not settle a disagreement between parties with different information or different interests. The bottleneck therefore moves from producing analysis to choosing among competing uses of intelligence and capital, and answering for the choice. Firms, managers, committees and contracts perform this role today. They were designed for a world in which analysis was scarce and a small group could review every plan that mattered. In our view, that world is ending.
A decision market turns disagreement into action
A prediction market prices a claim about the future. A decision market prices the same outcome under different actions and compares them. Futarchy, in Robin Hanson's formulation, lets that comparison choose the action: vote on values, bet on beliefs. The objective, and the way it is measured, remain institutional choices. The market answers the factual question of which action is expected to do better.
Consider an organization with a $10 million budget and two research programmes. AI systems propose experiments, estimate costs and forecast results for both. A decision market reduces all of that to one precise question: if funded, which programme will score higher on the agreed research evaluation by the agreed date? Anyone who believes the forecasts are wrong can take a position against them, and loses money if they are the one who is wrong. The comparison allocates the budget.
Only the funded programme produces a result. Trades on the other branch are called off, because the market can never observe what the unfunded programme would have done, and the contracts must say so. Choosing the evaluator and the score is part of the governance problem: no market makes a poor measure meaningful.
This is already in production. MetaDAO runs it on Solana. Each proposal opens a pass market and a fail market for three days. After the first day, a lagged time-weighted average of each price is recorded, and when the window closes the proposal executes if pass leads fail by the threshold set for it, with a lower bar for team proposals than for outside ones. The market does not advise a board. The price is the decision.
This turns a one-off vote into a loop: propose → compare → allocate → execute → evaluate. AI generates and investigates the options, the market gives informed disagreement an economic channel, the program carries out the choice, and the outcome informs the next decision.
AI makes the loop cheaper to run. Vitalik Buterin argues that AI participants make it economical to research small questions that would never attract enough human traders. His "info finance" starts from the fact a decision needs and designs a market to produce it, with the market as the engine and a trustworthy judgment outside it as the steering wheel. He also distinguishes AI that trades under a mechanism's rules from AI written into the rules as the judge, which invites attacks on the judge. We build for the first and stay out of the second.
Low-cost participation is no longer speculative. In July 2026 the Forecasting Research Institute reported that several AI systems are statistically indistinguishable from superforecasters on its ForecastBench questions. A forecaster that costs almost nothing per question is what a decision market needs to reach the thousands of small choices that no human crowd would price.
Decision markets do not create the money to act, remove liquidity risk or guarantee a good investment, and not every decision justifies one: the gain must exceed the cost of research, liquidity and delay. What a decision market does is let the best available information choose among feasible options. A grant programme, a treasury or an agent operator can begin with bounded decisions without handing over the whole organization.
An economy of intelligence needs markets to govern it
It is useful to separate two projects: creating a capable intelligence, and organizing an economy in which capable intelligences act. Here AGI means broadly general capability and ASI means capability that exceeds people across most important tasks. We use these as terms, not as claims that either threshold has been crossed.
Decision markets are not needed for the first project. A lab can build a superhuman system inside a centrally directed organization, and markets do not supply the learning architecture, human values or proof that an agent follows its mandate.
The second project is different, and it determines what superintelligence is worth. Researchers at Google DeepMind argue that general capability may first appear not in one system but in groups of agents with complementary skills, and that transactions between them should be governed by robust market mechanisms with auditability and oversight. We agree, and we name the mechanism. Once millions of independently controlled agents allocate real resources, review stops scaling. No committee can read every plan that superhuman planners produce, and no single model can be trusted to judge the rest, because a judge written into the system is a target for every agent it judges. Economies already coordinate what no central reviewer can see, and they do so with prices. A decision market is a price on a decision: any agent that knows better can put capital against a plan, whoever priced it wrong pays for it, and the record shows who priced it how.
The strongest objection is that AI makes the central reviewer possible after all. Erik Brynjolfsson and Zoë Hitzig argue that AI codifies tacit knowledge and processes it at scale, which weakens Hayek's case against central planning and favours larger, more centralized firms. Within a single firm we agree, and a firm that can review its own plans needs no market. An economy of independently controlled agents, however, is by construction one in which the consequential choices are made by parties no single reviewer controls. The question there is not whether a reviewer could compute the answer, but who would grant it the authority and who would check it. A market requires no grant of authority, and it corrects itself by paying whoever corrects it.
That is the thesis at full strength: decision markets are the institution that turns advanced AI into accountable economic agency. Whoever builds the intelligence, an economy of intelligences needs a way to compare competing plans, allocate bounded resources and expose every forecast to challenge. Other institutions still define the objectives and enforce the constraints.
| Function | What decision markets contribute | What must come from elsewhere |
|---|---|---|
| Choosing among plans | Conditional prices that make disagreement actionable | Clear alternatives, a sound decision rule, disclosure of material terms |
| Allocating resources | A signal that can inform or trigger a bounded allocation | Authorization, custody, budgets, contracts and execution controls |
| Evaluating agents | Incentives and records that separate useful forecasting from noise | Reliable outcomes, independence checks, resistance to metric gaming |
| Serving human purposes | What each action is expected to do to an agreed objective | The objective itself, legitimacy, rights, safety and technical alignment |
More intelligence does not remedy the ways markets fail. They fail if many agents repeat one source, if no one can take the corrective side, if gains outside the market pay more for manipulation than honest trading earns, or if the outcome measure can be gamed. Language models also fail together: across more than 350 of them, two models that both err agree on the wrong answer 60% of the time, and language-model pricing agents reach collusive prices without being instructed to. A market does not prevent these failures, but it pays whoever breaks a collusive or monocultural price, whereas a review board would inherit the same errors in silence.
The deepest objection is causal. A conditional price is not automatically the effect of the action, because which branch is chosen can itself carry information, so the spread mixes what the action does with what the choice reveals. Dynomight argues that no payout design fixes this. Hanson has conceded the bias in the cases Dynomight constructs and responds with design: announce when the decision is made, and occasionally decide at random so that some prices are conditional on a choice that reveals nothing. No production market does this yet. In our view it is the most important design change these markets still owe.
Intelligence therefore does not make decision markets work on their own. It makes the work around them indispensable: specification, participation, evidence and execution.
What the markets need that they do not supply
A market answers one question: given the information in the room, which branch is expected to do better? Everything around that question comes from elsewhere.
| The loop needs | Without it |
|---|---|
| Specification: clear options, checkable terms | People trade a proposal they have not read |
| Participation: independent information | Agents echo one source; no one takes the other side |
| Evidence: prices and outcomes, sourced and dated | Manipulation passes for conviction |
| Execution: proof the approved action happened | The market decides one thing, the treasury does another |
Together these are what the market needs and does not supply. The market prices a decision; it does not keep its record.
What Eidos builds
Eidos analyzes decision markets to explain what competing choices are expected to change, and how market conditions, trading behavior and available evidence shape that interpretation. The customer begins with a question: what does this market imply about one choice versus another, against which objective, and what remains unknown?
The conditional comparison is the starting point. A token-price difference is an implication about that token, not automatically a change in enterprise value or total organizational value. Any dollar-value conversion needs a justified valuation and supply basis, including changes specific to each branch. Where the mechanism prices a measured objective instead, the reading must stay in that objective’s terms.
The interpretation then needs evidence. Proposal text identifies what was promised; automated semantic checks describe the disclosed selection rule, including threshold and tie behavior. Branch dependencies and delivery enforcement need additional sourced review. Available depth and flip-cost scenarios describe price sensitivity under stated assumptions; the scenario is a lower bound, excluding fees and the cost of defending the price through the window. Order-flow metrics, VPIN and wallet concentration describe how trading formed the price. Sourced wallet histories can qualify named attack checks. Neither one-sided flow nor concentration proves manipulation.
The analytical record separates observed data, the domain model, instrument results and interpretations. Each calculation retains its method, evidence and time; incomplete inputs limit the affected result without suppressing unrelated observations. Sourced interpretations explain how the results bear on a decision and what remains unresolved. A no-match finding does not establish safety. An adjusted estimate or aggregate needs its own defined target, assumptions and validation against a baseline; individual measurements do not supply that validation.
The independent record makes the analysis inspectable: what it was for, what was known, what was done and what followed, each sourced and dated or withheld with its reason. Eidos is Greek for form. People read the evidence in the app and agents retrieve structured context through the API and MCP. Retaining source times and historical findings lets later readers challenge an interpretation against what was known then.
Passage, execution and outcomes answer different questions. The window outlook forecasts passage under the selection mechanism. Execution checks verify an approved action only where a supported rule and its evidence exist. Later outcomes report what happened after the decision. A market expectation, a selected proposal and an observed result do not by themselves establish decision quality or causal impact. The long-term thesis is that keeping these distinctions and accumulating comparable evidence will make markets easier to scrutinize as their use grows; that benefit must still be demonstrated.
Eidos does not trade, propose, decide, hold funds or run a market. It reports what markets imply and advises on specification and evidence, never what to approve, fund, trade or vote for. If a fund is ever formed, it will be a separate entity that reads the record only once it is published, on the same timestamp as every reader. Independence makes the analysis credible only while the controls can be shown. The Approach sets out implemented capabilities, evidence coverage and remaining work.
The evidence so far
The market is real but still small.
MetaDAO's founders page reported 22 companies and more than $45 million raised when we rechecked it on 10 September 2026. That is financing reported by the platform, summed across deals made on different terms. It shows that companies are raising through the venue, but not that they decide through it repeatedly. Our own read of the program addresses that.
On MetaDAO, after excluding eleven test organizations that never drew five wallets to a window, 22 organizations have run a market and 15 have run at least two proposals. The median window we could model drew 15 wallets, and depth is falling: the median was 16 wallets for windows in late 2025, 13.5 in the first half of 2026 and 8 for windows opened since July 2026, while the number of proposals rose.
New organizations keep arriving. Counting only complete months, 21 of them entered over eleven months: a launch rate of about 2.0 a month, between 1.4 and 2.5 on an 80% interval from a Poisson model of that pace. The rate counts organizations that put a decision to a market, the population above. It is not the founders page's count of companies that have raised, which is a different population, so it does not score a forecast made on that page's count.
The windows show where the decision forms. Most proposals pass: 32 of the 37 windows with enough prices. The price when a window opens is a worse guide to the result than simply guessing "pass". By the halfway point, the leading side matches the final decision about as often as that guess does, and at the close it matches almost every time, as expected, because the closing prices are what decide.
The practical implication is timing. A snapshot at the open says little. The verdict forms inside the window, and the useful reading is the continuous one taken before it closes, which is what the app records. The open question is whether the few proposals the market rejects were the right ones to reject, and only an outcome record kept over years can answer it.
The clearest case of the mechanism governing capital is enforcement rather than selection. In March 2026, holders of Ranger Finance's token proposed liquidating the project after alleging that the team had misrepresented the business before its raise. Our record of that window shows $693.44K traded over 246 trades from 60 wallets, and the proposal passed by 6.46 points over its threshold. About $5 million was returned to token holders with no committee involved. A passing market is not a finding of fact about the allegation. It is a binding decision about the treasury, which is what the market was asked.
The mechanism is not one venue. In February 2026, on Combinator, a market choosing among several options decided to redeem FairScale's entire treasury, on about $912,600 of reported volume against a project valued at about $348,000. Counting only real-money markets that reached a result and drew at least five traders, binding or advisory, at least 34 organizations have used one as of 19 October 2026, on five engines: MetaDAO, Combinator, Butter, Umia and Futarchy.fi.
Outside MetaDAO, the Uniswap Foundation's July 2025 pilot used a conditional funding market to award a $100,000 grant. It drew 29 forecasters and more than $70,000 in trading. The Foundation's own review identifies several problems, two of which matter here: insiders had more influence than expected, and forecasting complexity discouraged participation. The Foundation had announced a larger second round, but its case study reports only the first. It was a single allocation.
The only head-to-head comparison to date is Optimism's 2025 experiment, in which markets and the Grants Council each picked five projects from the same candidates, using play money. Forecasters predicted about $239 million of new value locked for their picks, and about $31 million arrived. Optimism's preliminary findings credit the markets' picks with about $32.5 million more growth than the council's, driven mostly by one outlier, and caution that three distinct projects a side and a volatile market rule out firm conclusions. GnosisDAO began a nine-month pilot in 2026 that shows conditional prices beside its votes, as advice voters are free to ignore. No one has yet shown, with real money and a fair comparison, that a market chooses better than a committee. That is the test this thesis most needs, and the one the record exists to make possible.
Investing in an economy of intelligence
Our goal is to make the path from an investment thesis to a funded decision, and from that decision to its outcome, open to examination. If intelligence expands what can be built, the institutions around capital must help choose what is worth building.
Investors need to understand the opportunity, the terms of their exposure and the conditions under which their thesis fails. A compelling account of technological progress is only a starting point. The work lies in connecting it to a specific use of capital, a plausible source of returns, and evidence that can change the decision.
That record must outlive the investment. What was funded? What did the organization promise? What changed? What did it deliver? People, and the agents they delegate to, need to carry those questions from research through allocation to review. Sources, assumptions and uncertainty should travel with every decision, so that a new reader can challenge it and its author can revise it.
Decision markets put competing judgments at stake. They work when participation is informed, outcomes are meaningful, and the approved choice is carried out. Building those conditions is the investment opportunity.
Capital is building intelligence. We are building the record that lets anyone challenge how capital was used and learn from what followed.
How the record can compound
Every market that decides something needs an independent referee: what the proposal promised, whether the market was manipulated, and whether the decision delivered. The chain does not make that position scarce. Its trades and account states can be bought and decoded by anyone, and a later reader can rebuild a market's history. What can compound is the referee's record, in stages that each have to be earned.
First, organizations supply what the chain does not hold: what a decision is for, who is bound by it, what was committed and the documents behind it, attributed to them and dated as they stood while the decision was open. They supply it because readers and agents cite the record. Second, each reviewed decision links what was promised to what followed, and those links become labeled cases the instruments are evaluated against at their original evidence cutoff, so reviews can improve with the number reviewed. Third, the review becomes part of how a decision is made: linked from a proposal, required by an organization's rules, read by an agent before it trades.
The mechanism matters. Contributions retain their sources and sharing scope; acceptance preserves the earlier record; an organization that stops answering leaves an empty field, never a carried-forward one; forecasts are evaluated against outcomes and baselines. Questions guide product priorities, while evidence tests correctness. Whether contributions and readers reinforce each other, and whether broader usage improves accuracy, are empirical questions, not assumed network effects.
What this means for the product today
Product bridge — 6 October 2026. The preceding conclusion states the long-term ambition. The immediate customer problem is narrower: a market has priced competing choices, and someone must explain what that comparison means and what supports it.
Eidos brings the proposal’s objective and conditional prices together with available evidence about depth, trading behavior and participation. Versioned instruments identify supported findings and unanswered questions. An organization can use that account to explain its market to stakeholders; a fund can inspect the basis of a conditional valuation; an agent can retrieve supported readings as structured context.
The record makes the explanation inspectable after the window closes. It preserves the sources behind the original reading while execution checks and later outcomes accumulate separately. That continuity connects the present service to the broader thesis: useful analysis now, and evidence for better comparisons over time. Prices and observations alone do not establish causal impact or decision quality.
The near-term test is whether customers find this interpretation useful enough to return and pay. Approach shows what the analysis contains and where its evidence stops. The dated forecasts and tests below retain their original terms.
What would prove us wrong
Three things must become visible this decade:
- AI participation lowers the cost of bringing useful, independent information into a market.
- Organizations continue to allocate through markets because their decisions improve after costs, not because the mechanism is new.
- Outcomes remain inspectable, so that a failure changes practice instead of disappearing behind the next proposal.
The thesis fails if conventional organizations absorb AI well enough that markets remain a niche. It fails if AI investment disappoints, funding tightens and the traders who would take the corrective side withdraw. It fails if windows keep drawing fifteen wallets or fewer while repeat use stalls, and if decision markets do not spread beyond a few dozen organizations; the register below dates that test. The first half of that test is already failing: windows opened since July 2026 drew a median of 8 wallets, and the depth forecast below gives one chance in five that the median recovers to 15 in the second half of 2027. Spread matters as much for the company as for the thesis: its own case needs about 500 market-governed organizations by 2030. And Eidos fails even if the markets succeed, if the organizations running them do not want an independent record, or will not supply the context its readers cite.
The record has its own test: if readers cannot act differently with the decision's form in hand than with the price alone, it is not worth keeping.
Trading volume, treasury size and capital raised cannot settle any of this. Two measures can: repeat use, and the quality of decisions after costs, measured against a committee or an AI-assisted team doing the same job. We have not yet measured decision quality. The register below tests the parts of the thesis to which we can attach a date.
The forecast register
Five dated questions test separate parts of the thesis and the business that rests on it. The probabilities are initial research judgments made on 29 September 2026 and 2 October 2026, not model outputs, and none has a track record.
- Adoption, 45%. By the end of 2027, will at least 60 distinct organizations, across at least three decision market engines, have had a real-money decision market decide or select one of their own actions? Decided by 31 December 2027; evidence by 31 March 2028. Registered · unresolved.
- Repeat use, 60%. By the end of 2027, will at least three different decision market engines each have an organization that has resolved two or more real-money decision markets on its own actions? Decided by 31 December 2027; evidence by 31 March 2028. Registered · unresolved.
- Depth, 20%. Across every decision market engine, will the median real-money decision market resolved in the second half of 2027 have drawn at least 15 distinct trading wallets? Decided by 31 December 2027; evidence by 31 March 2028. Registered · unresolved.
- Decision quality, 25%. By the end of 2028, will any organization or researcher publish a real-money, fair comparison showing how decisions selected by a decision market performed against decisions selected by a committee or vote on the same candidates? Decided by 31 December 2028; evidence by 31 March 2029. Registered · unresolved.
- Demand, 30%. By June 30, 2027, will at least one organization that decides by market have paid Eidos for a review of its own decisions, with the payment confirmed by a record the organization controls? Decided by 30 June 2027; evidence by 30 September 2027. Registered · unresolved.
Each resolved revision is scored with the Brier rule, (probability − outcome)², and missing evidence stays unresolved rather than counting as a no. The exact resolution rules and sources are in the forecast register.
Earlier work behind this revision
Before any forecast was registered, the register was rebuilt on 29 September 2026 so that each question tests decision markets across every engine rather than one venue or one organization. Three drafts were withdrawn unregistered: MetaDAO's reported company count, which measured fundraising rather than deciding; a second and third Uniswap Foundation round, which tested one organization; and the IEA's 2030 data-centre estimate, a neutral judgment that resolves in 2032. The cross-venue adoption question was revised from 100 organizations at 25% to 60 at 45%, because 100 was close to unreachable on the evidence and 60 is the pace the company's own case requires. The withdrawn drafts remain in the register. The registered rationales record the census as it stood that morning, at least 25 organizations with 19 on MetaDAO; later the same day, decoding Combinator's decisions in full and reading the program as of 29 September raised the floor to 30, with 20 on MetaDAO. The registered text is left as it was stamped.
The 18 September revision of this report built a ten-year stablecoin scenario model. When we replayed it against history, none of its 12 settings beat simply carrying the last value forward, so it no longer appears here. Its frozen inputs, model source, backtest source, default projection and validation results remain available as they were.