Graph based Decentralized Decision Making

A minimum degree of decentralization is table stakes for teams building in web3. However, decentralizing the governance of a product or protocol poses a significant amount of risk and difficulty. Under most commonly deployed incentive models, decentralization tends to slow the pace of decision making, it reduces the coherence of the overall vision and its associated work, it favors average opinions rather than surprising insights, and it diffuses responsibility so there’s no skin in the game. Anyone who has attempted to decentralize governance can attest to these challenges. This is a problem because it implies that decentralization — a core value of web3 — is at odds with rewarding high performing teams.
Competitive markets are frequently proffered as a suitable protocol to address this issue, they seem to balance an incentive for high performance with a degree of decentralization. Markets of course naturally arise in web3, with teams competing to attract paying users to their products and protocols. Unfortunately, a purely competitive market is also incompatible with web3 values — pure competition incentivizes closed source code, metered protocols, and private information, as well as inducing a generalized centralizing effect terminating in monopoly.
Instead, what’s needed is a protocol capable of spanning those two extremes.
This article concerns the creation of a protocol, tentatively called the Negation Game, that permits communities to allocate the capital in their treasury, and to permissionlessly take entrepreneurial risks. Instead of acting on the level of teams, as markets do, or on roles, as does, the primitive of interest to the Negation Game is the idea.
Players express their support for an idea in the Negation Game through staking a point. These points represent individual beliefs expressed in natural language, e.g. “The Earth is flat.” or a point can also be a policy that proposes to execute some action, e.g. “Build a rocketship.” or “Send 5 ETH to Alice.” In a mature protocol, players will stake points using the ERC-20 token of their choice.
The interface to interact with the Negation Game looks familiar to any user of Twitter, Discourse, or Discord: appearing as a series of posts that float across their feed, with prioritization given to beliefs that are considered “surprising” and “consequential”, as we will define below.
The ability to review the status of these policy conversations can be compared to the behavior enabled by proposal voting tools like , and is an uncomplicated design problem.
What’s unique to the protocol is the way in which it can aggregate many arguments and beliefs to inform a single conclusion, allowing for the cacophony of voices to be synthesized in a principled and credibly neutral way while still remaining inspectable and permitting nuanced discussion to reach the fore and impact the conclusion.
graph based interfaces make branching discussions easier to explore
The unit of account for these points is the credence, a measure of consensus and believability for each individual point after accounting for the support and relationships between all other relevant points. When the credence is high it’s a sign that it’s believed and supported in that moment. Credence can approximately be thought of as the amount of expected value at risk of slash if a statement is falsified.

The process by which a swarm anneals to an accurate credence value breaks down into three behaviors.
Players make and connect relevant points to one another, using the network’s existing beliefs to shift conclusions
Players dispute connections so that irrelevant information doesn’t flow
Players participate in unambiguous self-resolving prediction markets to ground the discussion in real-world observations
The benefit of this process as compared to other governance approaches is manifold. To name a few:
It incentivises the publication of private information and beliefs
It enables discussion to focus around the meaningful components of an argument
It allows reuse of values, preferences or beliefs that a community or group has already aligned on
It’s append-only, there’s no need for moderating deletion / inclusion as all questions of relevance or veracity can be disputed by adding or staking against a new point
As opposed to merely voting YES or NO on a proposal, it enables a conversation to ensue which itself educes a final proposal the community agrees with
It permits out-of-distribution ideas to gain currency because there is significant return for prescience
It’s epistemically sensitive: large token holders still get to have influence over the final conclusion, but whales may still lose if their predictions are inaccurate or if a competing position is more falsifiable
It’s bribery-tolerant: if you want a particular outcome your best move is to “bribe the protocol”, bribing someone else to stake on your behalf will allow them to charge a premium
The technical challenge this must overcome is to design an incentive landscape that gives even misaligned players a reason to identify information that could be harmful to their position. This is the research this project will fund. This looks like the process of identifying statements that if true would be incompatible with the existing statement, and providing additional influence for identifying observable statements. The protocol can then be equipped with the ability to marginalize over the positions held by players and to specially sample from highly trusted players in order to render preliminary conclusions.
A single proposal evolves like so:
Proposed and slated for execution By default, a belief that has any support is slated for execution as long as the time weighted average of support is large enough.
Counterpoints to reduce time weighted average support Any player may propose and stake a counterpoint, which puts the two points in competition with one another and both of their scores fall, reducing the time weighted average support over time and potentially unscheduling the execution. When two points are connected via a negation they behave like a prediction market, selling shares in each of them (or the possibility of both of them by way of the middle position). Thanks to this mechanism, a Negation Game enabled DAO can pay out of the treasury for activities that are believed by their epistemic graph, or if doubted idea entrepreneurs can buck the trend, buy shares in their own idea, and then later profit by upregulating its plausibility through making bets, making arguments, or running experiments.
Counter-counterpoints, objections, and crux identification Players follow their local incentive gradients to counter the arguments of other players and to argue that points are not relevant, thereby reducing their influence. This can continue until convergence, dynamically shifting the loci of the conversation to deeper root issues and recursively initiating new games, progressively iterating toward concrete claims that can be bet on. In this way a proposal isn’t purely executed by whales, but informed both by skin in the game and weight of evidence.
Identifying observables and betting Honest players interested in getting their way make concrete bets that their opponents agree are relevant and unlikely; in exchange they are granted additional influence. This is a mechanistic way to capture the aphorism, “a bet is a tax on bullshit.” The additional influence granted in this manner can play a role in shifting the conclusion of the conversation.
Resolution and slashing of leaves The protocol uses meta-markets, players’ prior moves, and graph topology to infer when a counterpoint-counterpoint pair is low ambiguity. In the case of a low ambiguity, concrete observable (what a platform would like Eigenlayer would call an ) the protocol can slash players if they don’t all reach consensus on the resolution. This gadget can be compared to the Schelling point systems used in Ethereum PoS and Kleros. These leaves of the graph then propagate their information back up, affecting their ancestor policies.

Let us consider three different kinds of uses for such a protocol:
Information value — “Under US law, tokens are classified as securities.”
Signal value — “This smart contract has no security flaws.”
Pragmatic value — “This project should be funded.”

Each of these is a distinct kind of value that the protocol provides.
Information value
A statement like, “Under US law, tokens are classified as securities.” is potentially useful clarity for someone starting a company that intends to issue tokens. Not only is a clear answer quite valuable, but even an ambiguous answer, e.g. “50% true”, can be quite informative, for example, as to the state of the legal environment that surrounds tokens, and can be informative to many parties for different reasons, like prospective business owners for whether to enter the space, or for regulators as to whether the regulations are clear and unburdensome.
This value of “making known what’s known” is among the primary benefits of dissensus protocols like Wikipedia and prediction markets — since it’s easy for anyone to dispute the existing result, and since there’s a benefit to doing so, you can accept the result of the disagreement as a decent prior and a starting point for further research.
Like prediction markets, this protocol can provide information value on various topics. What distinguishes this protocol from information markets on this front is the ability to ask more ambiguous and high level questions and still get credible results. A statement like, “Under US law, tokens are classified as securities.” would typically be considered undesirable to bet on by most market participants. This is because typical markets require a clear resolution criteria, and are subject to the whims of the judge. By contrast, the protocol makes it possible to meaningfully bet on a high level statement on the knowledge that its criteria and conclusion will later be settled sensibly due to its self-defining and self-resolving markets.
Signal value
The little acknowledged truth is that few prediction markets are actually information markets. Most prediction markets that garner attention and trades matter because they are used as signals — they are interacted with because they play a partial role in influencing some outcome to the world. The canonical example of such a market is, “Mitt Romney will win the US presidency.” Famously (if you run in the right circles) just before the presidential debates in the 2012 US presidential elections an unknown trader pushed the market higher by betting a significant amount that Mitt Romney would win the election. that this was done as a way to manipulate the market so as to make Romney appear more favored, which could be a sensible strategy if the trader believed that being perceived as favorable can cause you to become more favored.
If one thinks of these markets as sincere signals of epistemic value then this seems to be a flaw in the design. If, however, one considers markets to be a signal of a mixture of both epistemic and pragmatic signals then this property perhaps can be harnessed in quite a useful way.
Consider for example the statement, “This smart contract has no security flaws.” Doubtless, the authors of the contract would prefer that the answer is an unambiguous YES, especially if the credence of such a statement is shown to a user before they confirm their transaction. Due to its signal value, these authors would be willing to stake some amount in that statement so as to seed its credence. From an information market perspective they are “manipulating” the market, their bet is a mixture of both epistemic and pragmatic signals (they benefit from a particular market state). For an information market this is a problem, but we can make use of this behavior.
The author’s stake immediately benefits them by increasing the credence of the statement and thus the likelihood that users will confidently execute the contract. However, by staking they’ve also inadvertently created a bounty for security researchers to identify flaws in the contract, giving them a reason to pore over the code and potentially make a contrary bet. Furthermore, users of the contract have an interpretable signal of the contract’s trustworthiness, measured in both total credence (value at stake), as well as how long its been open to scrutiny and how much attention it garners. If the user just intends to make a testnet transaction, perhaps a low credence is acceptable. If a user is instead considering performing a transaction on main main the low credence value will cause them to think twice, “Its credence is 1.3 ETH and about to transact with a wallet that has more than 30 ETH. Yeah, not today.” In this way, a credence value can be to web3 transactions what star reviews were to web2 back when we believed we could manage Sybil problems.
Pragmatic value
Pragmatic value is the counterpart to signal value. With signal value, a participant may be willing to stake because they benefit externally from a particular state of the market, regardless of their belief in the truthfulness of a statement. In contrast, pragmatic value describes the a participant's willingness to stake based on their expectation of making a profit within the market by acquiring the asset. This could be due to their expectation of speculative value, as more people purchase the asset and cause its price to rise, or due to the possibilities of windfalls or future profit to be returned to those that hold the asset.
Pragmatic value is important for encoding information about the value of some outcome, and it’s the dominant driving factor in the valuation of speculative companies and projects, like startups. Yes, both Ben Graham and HODLers of GME owe their motivation to the category of pragmatic value. Without this pragmatic value — the amount returned to the investor in exchange for their investment — most markets would be barren wastelands, and so, like any market, the protocol provides returns to participants in the market as a reward for their help in estimating the size of the total outcome, the likelihood of success, and the expected popularity.
Pragmatic value within this protocol can be received for a variety of reasons that are unlike a traditional market. Yes, holding shares in a point can entitle a holder to a cut of profits generated by the organization. Yes, making more prudent bets will provide them with better returns. However, unlike in a prediction market, this protocol follows a “pass down” rule, which instead of paying out the counterparty to a bet, pays out the holders of the evidentiary counterpoints to a statement. This means that if for a market like, “Will it rain on Tuesday?” followed by it not raining on Tuesday, the YES holders will lose their holdings, but the value won’t merely be transferred to the NO holders, but rather to the set of counterpoints that proved informative in presciently bending the market toward NO (this could include NO holders, but due to nonlinearities in the influence of certain markets it might be a minority term). In fact, this could mean that a holder of a relevance statement of a counter-counter-counterpoint could be the biggest windfall recipient in the case it’s a highly surprising signal; perhaps in that case it encoded a signal that if seen would have been a flipped the sign to indicate rain with high confidence.
Want to print your doc?
This is not the way.
Try clicking the ⋯ next to your doc name or using a keyboard shortcut (
) instead.