This post will explore a novel mechanism we call “epistocracy,” a natural successor to futarchy, which facilitates autonomous consensus on the legitimacy of competing propositional statements for use in group decision making. We will study the space of proposed mechanisms which incentivize the discovery of the mutual relevance of a pair of claims in an epistocracy market; we call these Carroll mechanisms. Beginning with an initial set of relevant constraints and objectives,  we will employ game-theoretic techniques to identify feasible mechanisms and determine Pareto-optimal mechanisms within this set, thus setting the stage  for future implementation work. Our team includes an active Ethereum developer and a professor of Computer Science with expertise in game theory and information design; project results will inform future academic work as well as lead to practical Ethereum-based implementations, including for use in DAO governance and decentralized scientific funding.

Our proposed governance market greatly extends the primitives that are employed in current futarchy governance systems. As an introduction, a futarchy-based governance mechanism establishes a market for a governance proposition, e.g., “Send Alice 10 ETH to build the mobile app.” Governance participants can purchase shares in either the YES or the NO decision; YES purchasers are granted shares in the protocol if YES shares are more valuable than NO shares (and likewise for NO purchasers). Unfortunately, these types of Futarchy governance systems have substantial drawbacks:

• A governance proposition must be clearly defined before the market can begin trading. The process for discussing and defining these propositions typically happens “off-chain,” using peripheral channels external to the governance mechanism.

• Information relevant to the proposition (e.g., “Alice is not a software developer”) cannot be considered explicitly by the mechanism once the proposition is under consideration; it too must be discussed and litigated off-chain.

• Futarchic markets are exclusively forward looking. They cannot be used to retroactively consider the veracity of statements related to events that have already occurred.

• As with all markets, there is significant incentive to keep information private that might otherwise improve market decisions. This private information incentive undermines the ability to build markets that collaboratively run experiments for purposes of information gain (meaning that markets cannot be used for explore-exploit activities, including for uses like scientific research).

Our proposed governance mechanism, epistocracy, solves these challenges elegantly and aspires to be a successor to futarchy as a decentralized governance approach. What distinguishes epistocracy from futarchy is that epistocracy is model sensitive, meaning that not only is the probability of the event up for betting, but the story of why the event is likely to happen is also subject to market discovery. This is known as model elicitation, and is the key proposed contribution of this work. In practice, this means that unlike futarchy, epistocracy can start with a question like, “How should we grow users?” and perform policy search over the suggested actions, including recursively exploring their relative merits and demerits. Futurchy, meanwhile, can only begin once those options have been identified, and then only renders a YES or a NO decision, without any unpacking or transparency as to the reason why, nor the conditions under which it might make sense.

 Futarchy has long been heralded as a promising governance mechanism for use in the Ethereum ecosystem, including as far back as 2014 (https://blog.ethereum.org/2014/08/21/introduction-futarchy).  Futarchy has been touted for its possible ability to fix many of the game theoretic shortcomings with typical group decision making mechanisms. However, successful implementation and adoption of it within the Ethereum ecosystem has been limited, and the outcomes are mixed.

A key challenge to adopting futarchy is its unfamiliarity: futarchy is very different from typical group decision making processes. Typically, a small group makes decisions by discussing and aligning on the relevant attributes of the decision before finally jumping to the stage of evaluating it as a proposal and choosing to move forward. Futarchy’s process, by contrast, only begins after the proposal has been written. This means, of course, that by the time a successful proposal makes it to a plenary (the moment when a proposal is evaluated according to the decision mechanism), much of the difficult sensemaking, values alignment, and discussion has already taken place. The consequence, of course, is that far fewer voices were taken into account in informing the proposal, and in cases where finding agreement is difficult or the topic is nuanced, a proposal might not reach plenary at all.

 Implicit in the sensemaking and consensus building process is a host of important attributes, including:

1. Is this proposal aligned with our values and mission?

2. If we pass it, can we execute on it?

3. How will we evaluate our progress over time?

4. Is this the most informative experiment we could be running or would we benefit from collecting more information?

Futarchy, by its very nature, is only capable of evaluating the expected marginal value of an action with respect to some metric. But it’s not evaluating the relative value of the space of actions and possible action orderings. Nor does it provide value alignment and epistemic information — information that implies where a proposal is aligned with a group’s shared mission, and the places where it’s predicated on broadly shared and loadbearing beliefs which should be tested for soundness.

An interesting question to ask is whether it might be possible to directly integrate the game theoretic properties of the decision mechanism into the sensemaking process itself. Meaning that the entire process of constructing the proposal — including disputes, tentative experiments, value alignment, curiosity satiation, and conviction building — takes place not in peripheral channels, but directly under the influence of an incentive aligned decision mechanism. Such a mechanism might be superior in many ways, including by sampling from a broader plurality of voices as the proposal is constructed, and balancing the explore-exploit tradeoff that all organizations must manage.

Such an approach has been proposed and is being explored under the mononym epistocracy, which aspires to be a successor to futarchy as a decentralized governance approach. What distinguishes epistocracy from futarchy is that epistocracy is epistemically sensitive, meaning that the information content of the discussion is assumed to have value, not merely the pragmatics (in this context, pragmatics refer to anything that directly bolsters the value or sustainability of the entity, like gaining revenue, while epistemics refer to anything that allows the entity to determine what’s believable). In practice, this means that unlike futarchy, epistocracy can start with a question like, “How should we grow users?” and perform policy search over the suggested actions, including recursively exploring their relative merits and demerits. Futurchy, meanwhile, can only begin once those options have been identified, and then only renders a YES or a NO decision, without any unpacking or transparency as to the reason why, nor the conditions under which it might one day make sense. 

Epistocracy’s key innovation on top of traditional futarchy is through the use of disputable counterpositions. The easiest way to understand the usage and utility of disputable counterpositions is with an example:

Imagine there is a proposal like, “Send 10 ETH to Alice to build the mobile app.” In typical futarchy, this would be bet on and the result would merely render a YES or a NO. Let’s imagine that someone knows that Alice is not a developer, and that that might undermine the possible success of the project. Under futarchy, their best move is to credibly publicize that information with the intention to bring the market price down (ostensibly after buying more NO so they can also profit). But the challenges this faces are manifold:

1. How do they ensure the information disseminates quickly and broadly enough to inform market beliefs?

2. How can the market come to verify the legitimacy of that information, and do so in a way that doesn’t permit inauthentic arbitrage? E.g. through spreading fear, uncertainty, and doubt a nefarious participant may reap short-term profits at the detriment of the organization.

3. How can the market evaluate the importance of that information? I.e. if Alice is indeed an unskilled developer that’s quite concerning, unless of course she’s merely the manager of a team of developers, in which case it’s much less of an argument.

In a typical futarchic system, or just any market system generally, these whispers tend to be shared in peripheral channels, which would spur market participants to adjust their bets in accordance with the degree they believe it’s not yet baked in. In epistocracy, the information itself becomes another side of the market. Increased credence in this information can then decrease the resultant implied probability of the market by acting as the counterposition to the proposal. 

In other words, with this setup, the market participant finds that they have two affordances for decreasing the probability of the proposal passing. The first, as is typical, is buying NO shares in the proposal 10 ETH to Alice. Buying NO shares decreases the probability of the event passing. The second option is for the market participant to buy YES shares of Alice not dev. In this instance, for illustrative reasons, buying YES in Alice not dev has an identical effect to buying NO in 10 ETH to Alice.

To achieve this, we imagine initiating this design based on a typical Log Market Scoring Rule (LMSR) as would be used in an Automated Market Maker (AMM) (Hanson 2002):

Where Y refers to the number of YES shares that have been purchased of 10 ETH to Alice, and  N refers to the number of NO shares. As is typical, the price of the Y shares can be easily found by taking the derivative of the cost function. And therefore price is defined as:

To accommodate the impact of adding a market to estimate Alice not dev, we add a third share to the mix, just as we would with a categorical market.

Where B is the number of YES shares sold in the counterposition Alice not dev.

As we can see, this setup immediately grants us one of the desired properties of our epistocratic structure: namely, it immediately disseminates the new information into the market. It does so by A) immediately moving the market price, and thereby changing the probability of the proposal passing, and B) making the reason for the market move legible to other players, while permitting dispute of the veracity of the argument e.g. by buying up shares of NO Alice not dev. Already, we’re making progress toward our goal of fast dissemination, but we can’t yet expect that this will provide us with legitimacy and importance, which helps us understand why we need disputable counterpositions.

This naive design suffers from two immense flaws:

1. Under such a system, there’s no comparative advantage to acquiring YES Alice not dev compared to merely buying more NO 10 ETH to Alice. The two moves are (at this point) game-theoretically identical, which means that a player has no reason to prefer supporting the former over the latter (which is the behavior we’d like to incent so as to promote information dissemination).

2. What happens if someone adds an unimportant item as the counterposition? E.g. in the case where it turns out that Alice has a team of developers she contracts with, Alice not dev is no longer a compelling concern. It must be possible to dispute the importance of the counterposition. 

To address both, we parameterize the counterparty shares B with a weight q, which allows us to control the degree to which the B shares should have an effect on the cost (and therefore price) of Y.

It’s at this point that we permit q to become a disputable parameter which is governed by a secondary mechanism (to be named). As q rises through interaction with the secondary mechanism, the effect of B, Alice not dev, becomes stronger, decreasing the cost and price of Y. Inversely, as q falls the price of Y rises. 

Thus, with the addition of q, we’ve provided a parameter that allows us to achieve both of our stated goals: in the case that B is irrelevant to Y, the secondary mechanism can permit players to set q low, thereby minimizing B’s ultimate effect on the cost of Y. Inversely, when q > 1, our player has a reason to prefer purchasing B shares rather than N: they gain additional leverage over the probability of Y being enacted. (Note that there may be additional economic rather than pragmatic reasons why players want to buy B shares when q > 1, but the root motivation is considered to be gaining the ability to influence which proposal is selected, for a discounted price).

It’s worth noting that the introduction of q creates a host of new problems, which impose useful constraints on the behavior of the secondary mechanism that modifies q. As several examples:

1. Imagine a mechanism design where the value of q is directly controlled by another AMM; this would permit a form of price oracle manipulation attack where an agent could buy shares in the secondary mechanism, then buy shares in Y, then sell shares in the secondary, and then sell Y, thereby draining the liquidity supporting Y. The secondary mechanism must take this and other exploitation vectors into account.

2. If B has any outstanding shares, then increasing q necessarily decreases Y. But for a proponent of Y, this produces a pathological incentive — the market would prefer that a proponent of Y divulge their true reasoning, which might include assenting to the relevance of B in their decision (which they would do by participating in the secondary mechanism to signal B’s influence on Y). Unfortunately, any increase in q (which encodes the relevance of B) necessarily decreases the price of Y, disincentivising the honest player from revealing their information. This can be amended by providing a minor price premium to Y when a player holding Y makes a move that increases q.

Research funded by this grant will explore this constraint space and publish on mechanisms amenable to addressing these problems.

However, most of these are minor problems compared to the overarching problem that plagues all prediction markets: resolution. How do you decide how a market should close? For example, let’s once again imagine the example market that considers the claim Alice not dev, how do we consider that to be sufficiently answered that the shares can be paid out? How do we conclude, in other words, that indeed Alice is not a developer. Platforms like UMA and EigenLayer offer services for verifiable data which can be used to resolve low ambiguity questions, e.g. a question like, “What was the price of APPL on Thursday, Apr 10th at noon eastern?” This can be useful when needing to source a one-off fact, but when presented with a question like, “Is Alice a good developer?” those simple consensus-based mechanisms would be insufficient to the task.

Instead, epistocratic mechanisms take a different approach: they never resolve. Apart from select, low ambiguity questions that are amenable to consensus resolution (like market prices), shares in the system represent claims on the underlying assets (which come from the moves of traders), and they fluctuate constantly due to one of two reasons:

1. Due to the popularity of a share causing the price to be driven up or down, as with any market

2. Due to moves made in the secondary mechanisms which cause the connectivity (q) between two shares to change

The ambition, therefore, is to use the properties of the secondary mechanism to create stable attractor basins (think Hebbian learning × Schelling points) that inform players which shares are likely to remain durably popular. In this way, we can harness the massive capital accumulation behaviors of speculative assets, while directing the funds toward more productive ends. If we’re successful, epistocracy can be more than a novel governance mechanism; it can be a novel capital allocation mechanism for high ambiguity domains like scientific research.

The mechanism designer must be careful when designing such a system to ensure that the money used to set the parameters which influence these primary markets don’t simply rediscover the iron law that money gets its way. These sorts of designs will quickly become captured by the wealthiest players, and then, due to reflexivity, the wealthiest players would be able to use that power to spiral into more wealth. Therefore, in order to combat such capture there must be some noncapturable signal of trustworthiness that can be gleaned by the secondary mechanism so as to guide the setting of stable Schelling points.

The mechanism used for eliciting the sincere signal of trustworthiness in epistocracy is called epistemic leverage, and is where the system derives its name. The key idea of epistemic leverage is that a player who reveals information that would be surprising to them should also receive additional influence in the game. The system grants them this benefit because they are increasing the surface area of falsifiability of their preferred idea. This can perhaps be thought of as a mechanistic implementation of Karl Popper’s Theory of Falsifiability. But it can just as well be thought of as a tool for finding out that one is wrong (error detection) and listening to those who are willing to find out that they’re wrong (curiosity upregulation).

The key idea of epistemic leverage is quite simple: If a player already has shares in Y and they participate in the secondary mechanism to increase q (and thereby increase B’s influence on Y), the system upwardly adjusts Y by the amount that compensates the Y player in exchange for revealing their belief that B is important information (and thereby directing the network’s attention). The amount that the system awards in exchange for revealing this information is called reconciliation, and its size can vary. Complete reconciliation merely returns the player to their original level of wealth. But reconciliation need not merely be complete. It’s also possible to provide surplus reconciliation. Surplus reconciliation occurs when the network parameters are tuned so as to provide even more  Y shares to the Y player’s cause over and above the amount that would make them whole again. This provides a positive incentive for players to reveal private information about the reasons for their beliefs.

This state comes, however, with a certain risk: if players can gain free influence in exchange for saying “B is relevant” what ensures that they select a B that’s genuinely relevant and informative? In the absence of a steadying mechanism, players are incentivized to mark as many counterpositions as relevant so as to bolster their final score.

In order to combat this pathological behavior, there’s one final mechanism to introduce. This mechanism provides incentive closure to epistemic leverage, and imbues the mechanism with its epistemic properties. Given that a player using the secondary mechanism to increase q is making the statement, “This information is relevant”, this implies that as the shares sold in B rise, their willingness to change their mind about Y should also fall. What characterizes an insincere player is their unwillingness to change their mind as a result of changing information. Therefore, the final mechanistic affordance in the epistocratic stack taxes exactly that behavior, this affordance is called doubting.

When a player doubts another player, they are buying the right to A) reduce the free influence their opponent received, and B) begin to drain the currency that their opponent has risked, which they receive as payment. The rate of draining of the currency depends on the implied probability of B. If B is high (and q is high), then the rate of draining is also high, because ostensibly the player should have already changed their mind. Meanwhile, if B (or q) is low, the rate of draining is also low, because there’s not yet reason to change their mind. This payout is counterbalanced by a risk taken on by the doubter: if a player does change their mind (through an act called self-slashing) the doubter is slashed as well. The macro incentive this creates is one wherein the most trustworthy actors (the ones feared for their likelihood of changing their mind as a result of new incoming information) are the ones least doubted — and because they are least doubted they’re also the ones with the highest overall influence in the game. This engenders a generalized incentive to be known as the kind of player that changes their mind, which automatically selects for a host of good epistemic behaviors, including internal consistency, accurate forecasts, knowledge depth, and intellectual honesty.

The ambition of a completed mechanism set, therefore, is to use this mechanism of epistemic leverage to shift stable equilibria toward those shares that represent ideas that are consequential in their impact (connect to something with a high price and high liquidity) and durable in the face of opportunities for falsification. The dynamics by which we hope to achieve this is one of induced cascade — a design whereby what was once a safe position to hold suddenly becomes untenable because incoming information lower in the stack has caused the position to erode in its defensibility, thereby prompting self-slashes.

We imagine that the flow of events leading up to a cascade proceeds as follows:

1. A player takes up a position by buying shares of Y

2. Another player takes up the alternative position, buying into N shares

3. The Y buyer participates in the secondary mechanism which increases the value of q and provides a boost to the value of Y due to epistemic leverage

4. The N player believes the Y player won’t really change their mind as B rises, and so they doubt the Y player’s position

5. The value of B rises (as the market gains information), while q remains constant, and so the payout to the doubter starts to grow significantly.

6. Y starts to fall significantly in price, and so the Y player has two options:

   1. Exit their position by self-slashing, claiming a slightly better deal than they would have had the bleeding continued

   2. Fight back against the value of B by buying shares of the counterpositions of B and by using epistemic leverage against B

In this way, it’s possible to think about this as an iterated game, whereby the same mechanisms are used recursively to seek out deeper and more nuanced questions that underpin the issue. A useful analogy to think about this is as a dynamic bounty resolution mechanism that’s tied to a proposal. A player makes a proposal, they bump up its support by setting up standing bounties for more information (through the use of epistemic leverage), and then as people make progress in answering those questions either the player pays out those information bounties, or they dispute that they’ve been resolved; but the whole process is handled autonomously by the mechanism. If the network is unconvinced by their assessment, their stake will quickly drain away through doubting. This means that with each new information bounty “created” through epistemic leverage, there’s a ratcheting effect, causing the top level conversations to become sensitive to those deeper in the stack — and reciprocally, the ones deep in the stack stand to reap large returns due to the high liquidity of the shares found at more surface levels.

Epistocracy represents a significant evolutionary leap beyond traditional futarchy, transforming governance from binary evaluation of pre-formed proposals into dynamic collaborative sensemaking. By introducing disputable counterpositions and epistemic leverage, this mechanism makes the reasoning behind decisions as important as the decisions themselves.

The core innovation enables groups to begin with open-ended questions like "How should we grow users?" and collectively explore possible actions and their interdependencies, rather than merely voting YES or NO on predetermined propositions. The disputable counterposition system provides formal pathways for incorporating new information, while epistemic leverage incentivizes transparency by rewarding participants who reveal information that challenges their own positions.

The mechanism's non-resolving nature creates stable attractor basins that guide long-term capital allocation toward durably valuable propositions—particularly promising for high-ambiguity domains like scientific research where traditional resolution mechanisms fail.

Significant challenges remain around parameter manipulation, wealth concentration, and Schelling point stability. However, as this research progresses from theoretical foundations to practical Ethereum implementation, epistocracy has the potential to transform not just DAO governance, but our broader approach to collective decision-making in complex, uncertain environments. Success will be measured by real-world outcomes rather than theoretical elegance, positioning this work to contribute meaningfully to both academic understanding and the evolving landscape of decentralized governance.

Our proposed work falls into the broad area of decentralized governance and group decision-making, and there are a number of other active community projects which aspire to accomplish similar things. However, we believe that our proposed work is distinct from these other active projects in that it provides direct economic incentives for market participants to reveal what they know about the causal relationships between competing propositions. For the sake of completeness, here we briefly review several related projects.

Ethelo is a digital decision-making platform that helps groups navigate complex, multi-option decisions. Ethelo can explicitly model logical relationships between proposals as part of its decision structure. Decision designers or participants can define rules and constraints among options (for example, making two proposals mutually exclusive or requiring one as a prerequisite for another). However, Ethelo does not automatically detect and dispute contradictions in free-text proposals; the mutual relevance must be encoded in the decision setup. In other words, if the propositions “pay Alice to develop the app” and “Alice is not an app developer” are both on the table, a facilitator would need to input that dependency or conflict. Ethelo’s strength is in handling such interdependencies once defined, but it relies on humans to define the decision model. Thus, our proposed vision of epistocracy handles the problem of identifying logical relationships among competing claims in a way that is outside the scope of Ethelo.

One of the key possibilities enabled by epistocracy is that if the revelation incentives are properly designed, markets will output some type of aggregation of community opinions. Another system which approaches this goal directly is Pol.is, which allows anyone in the group to submit short statements on a topic, and all other participants can vote “Agree,” “Disagree,” or “Pass” on each statement. Behind the scenes, Pol.is uses statistical clustering to map how people’s opinions group together. One substantial difference between Pol.is and our vision of epistocracy is that Pol.is does not explicitly link statements to each other as “supporting” or “contradicting.” Each statement is standalone, and participants vote on it independently; any logical relationships between statements must be uncovered implicitly. In contrast, epistocracy will attack this problem directly.

Community Notes (formerly Birdwatch) is a crowd-sourced fact-checking system on X (formerly Twitter) that attaches context notes to tweets. Its goal is to allow the user community to add helpful context to potentially misleading posts​. Community Notes contains a clustering mechanism among fact checkers which aims to group checkers according to similar perspectives, and then publishes notes which receive approval across the spectrum of varying perspectives. The system thereby resists manipulation by any one faction: a coordinated brigade from one side won’t get their note shown unless some contributors from other viewpoints also find it helpful. This bears some loose similarity to our proposed system of epistocracy, but it is important to note that identifying logical contradictions between different governance proposals are outside the scope of Community Notes, which essentially handles one claim at a time, verifying or contextualizing it through crowd input. In addition, Community Notes is not explicitly a tool for group decision-making and thus solves a different problem than epistocracy.

For completeness, we briefly discuss several other related projects here. Discourse Graphs represent a framework and information model for mapping out arguments and ideas in a graph structure. The aim is to create a modular, composable graph of discourse that can be shared and updated collaboratively by many people. Discourse graphs represent relationships between ideas explicitly, but they are best thought of as a data structure rather than a governance mechanism: they can inherently represent conflict between claims, but have no mechanism for resolving this conflict.

Kialo is a centralized online platform specifically built for structured debate and argument mapping. The goal is to facilitate collaborative reasoning: by visually organizing points and counterpoints, Kialo helps users see the overall structure of complex issues at a glance. One potential shortcoming is that Kialo’s model assumes participants argue in good faith; if people game the impact ratings or spam with countless trivial points, the quality can degrade (though community moderation can address this). Epistocracy aims to address these incentive issues by design.