A decision market can compare two conditional futures with exquisite accuracy and still be answering the wrong question. The failure need not be a bad oracle, a broken average, or an irrational trader. It can begin one layer earlier, in the proposal that defines what pass and fail are allowed to mean.
That is the useful claim in a recent LessWrong threat model for asset futarchy. It is not an empirical finding that any live proposal used these attacks, and it is not a proof that futarchy fails. It is a list of candidate mechanisms: five ways a proposer might make the pass branch price above the fail branch without giving holders the unconditional value the comparison seems to promise. The right product response is therefore not to classify proposals from the essay. It is to turn each mechanism into a sourced, detector-pending question and name the evidence that would be required before answering it.
Five gates before the price
The first candidate is resistance-contingent delivery. A proposal can be valuable if delivered while quietly leaving delivery at the proposer's option. The proposer then compares the cost of doing the work with the cost of defending the pass price. Weak resistance makes non-delivery cheaper; strong resistance can trigger honest delivery. Either way, the book saw the value of the promised work, not an enforceable obligation to perform it. The review question sits in delivery enforcement: who must do what, by when, and what makes the obligation more than prose?
The second is conditional holder exit and corrective-supply failure. In a spot conditional market, selling the pass asset is not an abstract vote against the proposal. It is an agreement to exit the underlying asset in the world where the proposal passes. A holder may think a proposal mildly harmful and still prefer that diminished asset to cash. Non-holders may be unable to create the short exposure needed to correct the price; where synthetic shorts exist, capital still bears the risk that withheld good news makes the seller wrong. The candidate threat is a shortage of corrective participation, not evidence that every thin pass book is manipulated.
The third is proposal-convexity selection. A proposer can submit while a material detail remains unresolved. Passage then occurs disproportionately in worlds where favorable proposal-specific information arrives, so the pass market can correctly price a favorable conditional branch even when approving the unresolved proposal ex ante has negative expected value. This is the sharp edge of the conditional-versus-causal dispute: the market observes what value looks like given passage, while the decision maker wants to know what value passage causes. That gap is an assumption to disclose and test, not a causal effect the tape has already proved.
The fourth is countertrade deterrence through adverse selection. An underspecified proposal can make skepticism expensive. Missing counterparties, scope, fees, milestones, or enforcement terms may conceal a bad deal; they may also conceal a genuinely good one whose proposer plans to reveal the edge after buying pass exposure. A trader can believe the visible deal is probably bad and still decline to sell because the loss in the hidden-good state is larger. The market cannot qualify a specification it never received.
The fifth is fail-branch sabotage. A proposer with a credible ability to withdraw support, liquidity, or another dependency can make rejection worse. The market may then be locally right that paying is better than suffering the threat, while the organization would have preferred a standing rule that made such threats unprofitable before any proposal arrived. The missing property is branch independence: fail should describe life without the proposal, not life after retaliation for rejecting it.
These are gatekeeper mechanisms in the literal sense. Specification, delivery, participation access, branch independence, and the causal assumptions behind the comparison all sit upstream of the print. A reviewer can inspect them; a price cannot conjure terms that were omitted.
The repeated game pushes back
The one-shot examples give each proposer a cheap option and count little beyond the current proposal. Real organizations often repeat the game. A contractor who withholds delivery can lose later work. A known proposer who hides material terms can face a higher cost of capital. A liquidity provider who poisons the fail branch can destroy its own franchise, invite legal enforcement, or make holders precommit to reject the next demand. Reputation, recourse, and future order flow can turn an apparently cheap attack into an expensive one.
That counterargument matters, but it is not a universal defense. Pseudonymous proposers may have no franchise to preserve; one large extraction may dominate discounted future business; legal recourse may be unclear; and a threat that works off path may never have to be carried out. The empirical question is which repeated-game penalties actually bind in a given venue. A detector would need records of identities, prior commitments, delivery, sanctions, and exits — not merely a suggestive price shape.
One public experiment, one very small read
The MARS 4.0 research-selection experiment is a useful check on how little a decision result establishes. Geodesic opened conditional Manifold markets for research proposals, selected one project from each of three mentor groups, and locked the selected markets while their outcome questions matured. Reconstructing the closest choice from the locked public probabilities gives Goal Crystallisation an average 43.030% against Online Learning's 42.894% — a winning margin of approximately 0.136 percentage point. That is a close selection, not a precision estimate of research quality.
The reconstruction takes a simple unweighted mean across five answers. Manifold's public API preserves a resolved answer's lock-time value in resolutionProbability; an unresolved locked answer retains it in probability. Reading those fields from the two cited market records produces the averages above without relying on the rounded percentages shown on the market pages.
The first outcome shared by the three selected projects — a LessWrong post within six months that gained at least 50 upvotes within a month — has since resolved NO for all three. The next statement must be narrower than the tempting one. This does not show that the selected research failed: the market defined several later paper and conference outcomes, and research value is not identical to one publication threshold. It does not show that futarchy failed: we do not observe the unselected projects under selection, so there is no counterfactual outcome against which to score the choice. Nor can the market's locked probabilities be treated as a causal estimate of what selection did. The tape records a close decision followed by three NO resolutions on one shared KPI. Anything larger is invention.
Conditional is not yet causal — and that is not a eulogy
The literature is openly divided. One line argues that decision markets necessarily reward conditional correlations rather than causal effects when their prices help choose the action. A published response argues that explicit randomization — occasionally taking a decision independently of the market and cancelling most contracts — can recover the causal quantity, at the cost of more committed liquidity and an execution rule real institutions may resist. The gatekeeper threat model adds proposal-controlled semantics to the same argument. None of those papers settles how large the effect is in a live asset futarchy with repeat players, enforceable terms, and observable delivery.
So the honest state is neither approval nor disproof. Chromatic carries that boundary into the versioned vector library, where these five mechanisms remain documented and detector-pending. The same product surface publishes the six color-team contracts and the evidence each definition would need. Documented vectors are not assessed and do not classify proposals. Futarchy remains a mechanism for buying conditional information and acting on it. The new work is to state where that conditional comparison relies on causal assumptions, make proposal terms enforceable enough to price, and retain the outcomes and falsifiers that could turn a candidate mechanism into evidence. Until those detectors exist, these five threats stay documented and unclassified. A market can reveal what traders believe about the branches it was given. The product's job begins by checking that those were the branches the organization meant to choose between.