In 2003 Robin Hanson worked out what it costs to buy an answer. A sponsor who wants a question resolved posts an automated market maker and lets anyone trade against it. Under the logarithmic rule, the sponsor's worst-case loss is fixed in advance — b·log n for n outcomes — and the expected loss is exactly b times the entropy the traders remove from the sponsor's prior. The subsidy is not a cost of running the market. It is the purchase: dollars in, bits out.
That one result organises everything this desk covers. A prediction market prices an outcome. A conditional market prices an outcome inside a branch that may never exist — the trades are "called off," in Hanson's phrase, if the condition fails, a structure Paradigm's Multiverse Finance generalises into whole parallel balance sheets. A decision market runs two conditional books against each other and executes whichever future prices higher. An impact market runs the tape backwards: fund the work now, and let a later buyer price what it turned out to be worth. One family, one constraint. What separates the legs is who consumes the answer, and who pays for it.
The eighteen months to August 2026 ran the experiment at scale. Betting on events became a nine-figure-a-day industry. Betting to decide stayed a rounding error on the same rails. Betting on what already happened — the impact leg — was quietly wound down by its own sponsors. The gap is not accuracy, legality, or ideology. It is that only one leg found someone else to pay for its answers.
What the theory settled
Three results carry the literature, and they are older than most of the venues.
Accuracy was never the problem. Across 964 national polls and five US presidential cycles, the Iowa Electronic Markets sat closer to the final two-party vote than the polls 74% of the time — mean absolute error 1.82 points against 3.37 — and the edge was largest far from the election, 2.65 against 4.49 points beyond a hundred days out. Inside firms, Cowgill and Zitzewitz found markets at Google, Ford, and a third company beat official expert forecasts by up to a 25% reduction in squared error — and every one of those programmes was shut down anyway, by politics, cost, and lost sponsors, never by inaccuracy. Manipulation, the standing objection, mostly fails in the field: attacks on the IEM and the historical Wall Street election books moved prices only briefly before informed money reverted them, and Hanson and Oprea showed a manipulator with a known agenda is a subsidy — noise that pays informed traders to show up. The caveat that matters for this desk: that defence assumes elastic informed capital, and it fails outright when the payoff the manipulator wants is the decision itself, not the price. In a thin book, nobody is holding the other side of the trade.
The decision is where the incentives bend. A market asked "what is the firm worth if we do X?" prices a conditional expectation, and the condition is not random — it is chosen, partly on the market's own signal. Hanson named the failure himself: when deciders hold private information, conditional prices measure selection, not treatment. Othman and Sandholm sharpened it to a theorem: under the natural rule — do whatever prices best — a trader can always profit by exaggerating the worse option, and no scoring rule fixes it. Chen, Kash, Ruberry, and Shnayder proved the honest version exists but carries a strange tax: the decision maker must "always risk taking every available action" — commit to sometimes executing the option the market priced worse. No production system pays that tax, which means every live decision market runs on incentives that are approximately, not exactly, straight.
The escape hatch is equity. Oesterheld and Conitzer showed that a principal who insists on deterministically following an expert's recommendation can honestly elicit only two things — the best action and its expected value — and only with payment schemes that "give or sell the expert shares in the principal's project." That is a single-expert theorem, not a market result; nobody has extended the guarantee to many strategic traders. But the instrument it names is the one the surviving futarchy built: MetaDAO's traders do not hold abstract claims on a welfare metric, they hold conditional claims on the organisation's own token — exposure, not scoring points — and the market settles into the asset itself, with no external oracle adjudicating the outcome. Theory spent a decade proving scoring rules bend at the decision; practice converged, independently, on the same instrument the single-expert theory reaches for.
The same split explains the graveyard. Advisory markets — corporate programmes, GnosisDAO's ignorable signal books — die of their hosts, because an advisory answer embarrasses whoever it grades and binds no one. Hanson's own fix for selection bias is commitment: "put market estimates directly in control of decisions." Binding is not a bug in futarchy. It is the only version the theory endorses and the only version still alive.
The tape, August 2026
| Leg | The answer | Who pays for it | State |
|---|---|---|---|
| Betting on | What will happen | Gamblers, voluntarily | ~$214B YTD across two venues |
| Betting to decide | What we should do | The org, explicitly | ~100 binding proposals, one venue class |
| Betting on what happened | What it was worth | Nobody found | Retreating to grants and registries |
Betting on is no longer a niche. Tracker-reported 2026 volume through mid-August: roughly $155B on Kalshi and $59B on Polymarket — against about $40B for both combined in all of 2025. Kalshi raised at a $22B valuation and was reported in June to be seeking roughly twice that; ICE, the NYSE's parent, has put about $2B into Polymarket across two rounds; Robinhood's event-contracts revenue reached $156M in Q2 2026, more than its equities business. Two caveats belong next to every one of those numbers. The mix is a sportsbook: 89% of Kalshi's 2025 fee revenue came from sports, and contracts with any decision relevance are a sliver of throughput. And the notional is inflated: Columbia researchers put wash trading at roughly a quarter of Polymarket's historical volume.
The epistemic layer is real anyway — the Federal Reserve's own working paper finds Kalshi's macro markets well calibrated and its CPI prices beating the Bloomberg consensus — but it is a by-product, cross-subsidised by entertainment flow. And it sits on a resolution layer that keeps buckling at size: a $7M Ukraine-minerals market flipped by a whale casting about a quarter of the oracle's votes; a $160M+ market adjudicated by an oracle token whose fully diluted value was smaller than the market it settled. The fixes were whitelisted proposers and in-house discretion — the decentralised truth machine retreats to a referee exactly where the money gets large.
Betting to decide is four or more orders of magnitude smaller, and the desk's own tape says so plainly: April printed $2.84M of decision-market volume, July printed $5.86K, and a single large proposal is most of any good month. But the record is not volume, it is verdicts — roughly a hundred binding proposals across a dozen-plus organisations since late 2023. A market refused 62% of a treasury for $4,623, fired a team and returned ~$5M to holders, and renegotiated a $2.7M proposal mid-window. Where the mechanism was oversubscribed is not governance but capital formation — Umbra's ICO drew $155M in commitments against a $3M cap because the market is the covenant on the treasury. The 2025 cohort raised $96M on that pitch. The covenant already sold; the books are what's thin.
The conditional-funding experiments sit between the legs, and they returned data. Optimism ran futarchy beside its own Grants Council in 2025: 430 forecasters, ~1M OP allocated, and after 84 days the market-selected projects had added roughly $63.5M of TVL against $31M for the council's picks — with honest caveats attached (play money, forecasts that overshot ~8x, a USD-denominated metric polluted by ETH's price). The Uniswap Foundation ran the real-money version: $900K of grants steered by conditional markets, 29 forecasters, about $70K of volume, and a first-round winner that beat its forecast TVL by 11.8%. Small, priced, and — unusually for governance experiments — scored.
Betting on what happened failed by its sponsors' own grading. Optimism earmarked 850M OP for retroactive funding and deployed roughly 60M across six rounds before narrowing and cutting the programme; its retrospective concedes "we don't (yet) have the data to show that retroactive funding produces superior outcomes." Scott Alexander graded his impact-certificate experiment 4/10; Manifund's post-mortem concedes "we still haven't found a use case where certs led to better funding decisions"; hypercerts pivoted from a market standard to an open-records registry. The decade of impact bonds off-chain tells the same story at ~$421M cumulative — less than one mid-sized fund.
The price-of-an-answer frame says why. An impact certificate is a conditional claim with no natural resolution: attribution is genuinely contested — GiveWell's own counterfactual adjustments span 0.8x to 2x on the same intervention — so the claim settles only by fiat, and a fiat-settled certificate prices the judge, not the impact. No underwriter showed up for that asset. Hanson's subsidy bound is finite because the question ends; here the bill is undefined, so nobody paid it, so there was no market. Retroactive funding is not an early market. It is a grants programme with a delay.
Liquidity is the mechanism
Look at who actually pays for depth on each venue, because every one of them pays differently, and none of them pretends the answer is free.
Polymarket meters the subsidy per minute: resting quotes earn ((v−s)/v)²·b — quadratic in closeness to the mid, linear in size — and a one-sided quote earns exactly a third of two-sided credit. Takers pay fees that peak at even odds. Kalshi files its market-maker contracts with the CFTC: continuous two-sided quotes, maximum spreads, minimum sizes, term-limited. And the GWU study of 46,000 Kalshi contracts shows who funds all of it: takers lose about 20% on average, buyers of sub-10¢ lottery tickets lose over 60%, makers out-earn. The gamblers pay for the epistemics, at scale, voluntarily. That is the entire business model of the betting-on leg, and it is why it did not need a theory of who buys the answer. The control case is Augur: real demand to bet, nobody paid to quote, and the books emptied within a year of its 2018 launch.
A decision market cannot free-ride on that flow. Its book exists for ~72 hours, splits into two conditional branches, and extinguishes at resolution — the worst possible product for an organic market maker. So the venues that work pay explicitly, in structure. MetaDAO serialises to one live proposal per DAO, moves half the spot liquidity into the conditional pair for the duration, and gates the verdict behind a lagging, per-update-capped TWAP, so moving the answer means holding a distorted price against depth for days. The moat is time multiplied by depth, and the desk prints depth next to every pass edge because thin is sometimes consensus and sometimes cheap. Umbra's defence held with ~$52K in the pools on a day $1.5M was in play. It worked, and it was not deep — the covenant is priced, the standing books are not yet.
Butter's design documents state the desk's thesis better than the desk did: DAO-funded liquidity rewards are "payment for the information the DAO gains." That sentence is the category's income statement. The LMSR's b, Polymarket's reward pool, Kalshi's market-maker schedule, MetaDAO's donated pool share — the same line item under four accounting systems.
The research frontier is about making that line item smaller. LPs in a constant-product pool bleed to arbitrage at σ²/8 of pool value per unit time, and outcome tokens are the worst case — volatility explodes near expiry and at extreme prices. Moallemi and Robinson's pm-AMM flattens the bleed into a constant rate and prescribes withdrawing liquidity on a √((T−t)/T) schedule: liquidity for a dying asset should be scheduled, not standing.
Hanson's combinatorial result points the other direction — a full exponential outcome space costs no more subsidy than its marginal questions, entropy being subadditive, though exact pricing is #P-hard. And Multiverse Finance sketches where the primitive generalises — conditional collateral, conditional lending, whole verses of finance — with the honest parenthetical "liquidity issues aside." That parenthetical is this essay's subject.
Where it's going
Ranked by the weight of evidence, not by preference.
The sportsbook funds the epistemics. The base case, already priced. Event-contract exchanges converge with regulated gambling, and the answer factory rides along as a data product — ICE distributing Polymarket feeds, the Fed citing Kalshi, small businesses starting to hedge tariffs and weather on rails a World Cup paid for. Decision-relevant liquidity keeps compounding as a passenger on entertainment volume.
Decision markets grow where the answer has a named buyer. The two proven buyers are launch cohorts purchasing a covenant — capital formation, the one oversubscribed demand this mechanism has ever seen — and grants programmes buying allocation signal at five or six figures a round. Both are bounded by explicit subsidy budgets, not by belief, which makes the growth path legible: watch treasuries and foundations line-item the information spend. A useful tell will be the first DAO that budgets proposal liquidity the way it budgets audits.
AI cuts the price of the answer. More than 30% of Polymarket wallets already sit behind agents; Olas counts 18.75M agent transactions on Gnosis. The accuracy is not there yet — superforecasters still beat the best model on ForecastBench (Brier 0.096 against 0.122), and six frontier models each lost double digits trading real money on Kalshi over 57 days, though the same models roughly broke even on Polymarket. But the economics do not require AI to beat the crowd. They require the marginal informed opinion to get cheaper, because every dollar off the cost of forming a view comes off the subsidy a sponsor must post to get a thin question answered. That is Vitalik's info-finance claim made mechanical: micro-markets become viable when machine attention undercuts the subsidy.
Impact pricing returns as an algorithm, not an asset. Deep funding — AI-proposed weights over a dependency graph, human juries spot-checking — is allocation without underwriting, which is what the failed certificate experiments were actually demanding all along. Expect impact measurement markets attached to conditional funding (pay for the forecast, not the claim) rather than a certificates revival.
The wall that stays up. A market on a named company's decision — fire the CEO, do the merger — is, on its face, a swap on an event "relating to a single issuer": SEC territory, and no registered venue lists one on a merger or a management change. Corporate futarchy, the oldest dream in this literature, remains legally homeless in the US even as event contracts boom around it. The mechanism's near-term future is organisations constituted under it, not institutions converted to it.
Open problems
Four things would move this category more than any headline.
- A market-maker contract for a 72-hour book. Kalshi files DMM terms for standing markets; nobody has written the equivalent for a book that lives three days and dies. The pm-AMM's scheduled-liquidity result is the right starting point — decision markets are the extreme case of the dying asset.
- Shared-inventory AMMs for conditional pairs. Pass and fail books are near-perfectly correlated and currently fragment depth; a design that lets one inventory quote both branches would roughly halve the subsidy.
- Entropy accounting. The cost of an answer has been computable since 2003 — prior, outcomes, b. A treasury should be able to put "this decision is worth answering at $X" in a budget line and have the market maker enforce it. Nobody does this. It is the most shovel-ready idea in the category.
- Resolution that scales with notional. An oracle whose security budget is smaller than a single market it settles is an unpriced attack waiting for size. Settling into the asset itself — the futarchy design — is the only known answer, and it only works when the organisation is the question.
The desk's read: the answer factory got built, and gamblers are paying its power bill. The instruments that consume answers — decision books wired to execution — stay small until their sponsors treat the subsidy as what Hanson priced it to be in 2003: not overhead, but the purchase. Some already do. They are the ones this desk indexes.