CORE: Consensus Or Random Exclusion Endowment, Fair Exclusion, and the Scope of Consensus Kōmyō (Hiveism) and Claude (Anthropic) 2026-11-15 Abstract We present Consensus Or Random Exclusion (CORE), a procedure for groups that have to agree on a single course of action. The participants deliberate to- ward unanimous agreement, and at any time any of them may call for one par- ticipant, drawn uniformly at random and possibly the caller, to be removed. If removals continue, the procedure ends in random dictatorship, so that each par- ticipant’s fallback is a fair lottery over the participants’ ideals, which we call their endowment. A call never changes anyone’s expected endowment, and it has the same effect whoever makes it. The lottery gives weight only to first choices, so a compromise that everyone values and nobody ranks first can be reached only by agreement. CORE is therefore not an algorithm that computes a result from ballots but a frame for bargaining: deliberation, which we leave without formal rules on purpose, is the search for such outcomes and the bargaining over how to share what they gain. No participant and no group can force an outcome. All that anyone can force is the step from deliberation to the draw. When smaller groups can agree after a removal, an agreement survives the refusal of a few, but value can shift toward or away from those removed, and we state the guarantees that hold regardless: every participant keeps a 1/𝑛 chance of their ideal, and any 𝑚 participants keep a chance of 𝑚/𝑛 for an option they share. A decision cov- ers only what its participants brought into it. We discuss who should take part, interpret the mechanism as a veil of uncertainty over who will be decisive, relate it to the consensus methods of Heitzig and Simmons, and sketch extensions to sortition and delegation. 1 Contents 1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3 Building intuition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3 Where it could be used . . . . . . . . . . . . . . . . . . . . . . . . . . 5 Outline of the paper . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6 2. The mechanism . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7 3. The endowment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9 4. The call for random exclusion . . . . . . . . . . . . . . . . . . . . . . . . 9 5. Consensus and the fallback . . . . . . . . . . . . . . . . . . . . . . . . . 11 6. Where the gains from agreement lie . . . . . . . . . . . . . . . . . . . . 12 7. When smaller groups agree . . . . . . . . . . . . . . . . . . . . . . . . . 14 7.1 Benchmark and holding can come apart . . . . . . . . . . . . . . . 14 7.2 The effect can also run the other way . . . . . . . . . . . . . . . . 16 7.3 Free-riding and hold-up . . . . . . . . . . . . . . . . . . . . . . . . 16 7.4 Two variants . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18 7.5 What holds in both variants . . . . . . . . . . . . . . . . . . . . . 19 8. Participation and the reach of a decision . . . . . . . . . . . . . . . . . . 21 8.1 Why people take part . . . . . . . . . . . . . . . . . . . . . . . . . 21 8.2 What a decision covers . . . . . . . . . . . . . . . . . . . . . . . . 23 8.3 Changing what one does is not taking . . . . . . . . . . . . . . . . 23 8.4 Effects beyond the group . . . . . . . . . . . . . . . . . . . . . . . 24 8.5 What a majority can do . . . . . . . . . . . . . . . . . . . . . . . . 24 9. Strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25 9.1 What the impossibility theorems establish . . . . . . . . . . . . . 25 9.2 Incomplete information . . . . . . . . . . . . . . . . . . . . . . . . 26 9.3 Where the remainder lies . . . . . . . . . . . . . . . . . . . . . . . 27 9.4 Dividing the gains . . . . . . . . . . . . . . . . . . . . . . . . . . . 27 9.5 Why deliberation is left undefined . . . . . . . . . . . . . . . . . . 29 10. Scope . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31 11. Interpretation: a veil of uncertainty . . . . . . . . . . . . . . . . . . . . 33 11.1 A veil over decisiveness . . . . . . . . . . . . . . . . . . . . . . . 33 11.2 The prisoner’s dilemma . . . . . . . . . . . . . . . . . . . . . . . 34 11.3 The veil in the two variants . . . . . . . . . . . . . . . . . . . . . 34 11.4 The veil and the choice of rule . . . . . . . . . . . . . . . . . . . 35 11.5 Safe and worthwhile . . . . . . . . . . . . . . . . . . . . . . . . . 35 12. Relation to existing work . . . . . . . . . . . . . . . . . . . . . . . . . . 35 13. Outlook . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36 13.1 Sortition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36 13.2 Delegation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37 13.3 Adoption . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39 14. Open problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39 15. Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42 2 1. Introduction Suppose a group has to agree on a single course of action, and its members want dif- ferent things. The three familiar ways of settling the matter each fall short. Majority rule lets the larger side impose its will on the smaller. Strict unanimity looks like the fair alternative, but it lets any single member block a decision indefinitely, and block- ing keeps things as they are. Whoever benefits from the way things are can therefore block everyone else. Unanimity does not avoid imposition. It hands the power to impose to the status quo. A random lot treats everyone alike, but it cannot find an agreement even when one exists. In this paper we present Consensus Or Random Exclusion (CORE), a fourth way of reaching a decision, which combines the search for unanimous agreement with the fairness of a lottery. It consists of two moves: 1. The group deliberates, looking for a proposal that everyone accepts. 2. At any point, any participant may call for one participant to be removed from the deliberation. The person removed is drawn uniformly at random from ev- eryone present, including the one who called. The remaining participants then continue with step 1. This is the whole mechanism. The rest of the paper explains why it works, what exactly it guarantees, and where its guarantees end. The first move has no further rules. It is not a vote, and nothing fixes how proposals are made or in what order they are taken up. CORE is therefore not a voting procedure that computes a result from ballots. It is a frame for bargaining. The frame fixes two things: when a decision has been reached, and what happens if none is. Within it the participants search for an agreement and bargain over its terms, and all of them have the same fallback and the same moves. The absence of further rules is not an omission. Section 9.5 argues that any formal rule for deliberation would build an exploitable advantage into the procedure and would confine the search to options someone had thought of before it began. Deliberation has to be left open. Building intuition It helps to start from the case in which agreement never comes. If participants keep calling for removals, the group shrinks one member at a time until a single participant is left, and that participant’s preferred outcome is enacted. Because every removal is a uniform draw, each participant is equally likely to be the one left at the end (equivalent to random ballot). As long as no agreement is reached, then, each of the 𝑘 participants present has a 1/𝑘 chance of being the one who decides. What a participant holds in this situation is a fair lottery over everyone’s preferred outcomes, and we call its value to them their endowment . It rests on a probability of being decisive, not on a share of a vote, and it is the fair benchmark against which every agreement is measured. Two properties of the endowment make the mechanism work. The first is that calling 3 for a removal does not change its value. A participant who calls may survive into a smaller group, where their chance rises from 1/𝑘 to 1/(𝑘 − 1) , or may be the one removed. In expectation these two possibilities cancel exactly, and they cancel for every other participant as well (Theorem 1). This is also the reason the caller has to stand in their own draw. If callers were exempt, calling would become profitable, and the participants furthest from the rest of the group would call constantly (Corollary 1.3). The second property is that the endowment is usually a poor deal. A lottery over everyone’s preferred outcomes is a gamble between extremes. Most of its weight falls on outcomes that someone else wants most, and those are often much worse for you than a reasonable compromise would be. There is therefore usually room for an outcome that every participant prefers to their own endowment at the same time. This gap is what drives the mechanism. Deliberation searches for an outcome that beats the lottery for everyone at once, and unanimous agreement is the exchange in which each participant gives up their chance of getting everything in return for something they prefer to that chance. The lottery cannot find such outcomes, because it counts only first choices. An outcome receives exactly the share of participants who rank it first, however much everyone else may value it, so a compromise that nobody ranks first has no chance at all (Theorem 3). Whatever a group gains by agreeing, it gains by giving some outcome more weight than its share of first choices, and only deliberation can do that. This is why deliberation is the substance of the procedure. It is also why the mechanism has an honest limit: where the disagreement is purely about dividing something fixed and no compromise exists, there is nothing to find, and CORE returns the fair lottery (Proposition 1). This gap is also the reason anyone should take part. The fairness results that follow describe how the fallback is shared, and it is easy to read them as if the point of the mechanism were only to stop anyone from gaining an unfair advantage. But nobody joins a decision merely to avoid being cheated. People join because coming together can leave everyone better off than they would be apart. CORE is designed so that every participant can look for this gain from the same position, with the same fallback and the same moves. The guarantees are the floor. The reason to take part is what can be gained above it (Section 8.1). Usually more than one outcome beats the lottery for everyone, and the participants then also have to settle which one they adopt. That part is a bargaining problem. Everyone shares an interest in reaching some agreement, and they differ over which, so the gain from agreement is collective while its division is contested. Section 9.4 discusses how this bargaining can converge, and why no rule can settle it in advance. One complication should be met early. The picture so far assumes that once removals begin, they continue to the end. But a smaller group is still a group, and after a re- moval it may find an agreement of its own. A removed participant then no longer receives the lottery over the remaining preferred outcomes. They receive whatever the remaining participants agree on. That may be worse for them than the lottery, if the others share a compromise they dislike, or better, if the compromise suits them. 4 Section 7 works this out and distinguishes two variants of the mechanism. In the sequential variant, which is the one described above, smaller groups can find agree- ments the lottery would miss, and those agreements can shift value toward or away from the participants who were removed. In the one-shot variant, a stalled deliber- ation goes directly to the lottery, so that the endowment is exactly what everyone holds. The variants also differ in what a single refusal can do. In the one-shot variant, one participant who refuses sends everyone to the lottery, and in a large group it is al- most certain that someone will. In the sequential variant, a participant who refuses is likely to be removed, and the others can still agree. For this reason we recommend the sequential variant for all but small groups. Both variants keep a core of guarantees. With these ideas in place, three consequences can be stated before the formal devel- opment. The first is that the removals are not meant to happen. The uniform lottery over the participants’ preferred outcomes is itself a proposal the group can adopt, and it leaves everyone exactly at their endowment. The possibility of removal does its work by being available, since it fixes what any agreement has to beat (Theorem 2). The second is that no participant holds a special instrument. The only move anyone can make alone is the call, and a call triggers the same random event no matter who makes it. A majority has no lever that a minority lacks. Every participant can secure a 1/𝑛 chance that their preferred outcome is enacted, and any 𝑚 participants acting together can secure a chance of 𝑚/𝑛 for an option they share (Section 7). The third concerns the reach of a decision. A decision reached under CORE is valid only among those who took part in it, and only over what they brought into it. It cannot authorize anything beyond that. People outside the group who are affected by what the group does keep their claims, and the way to settle those claims is a wider agreement that includes them (Section 8). We deliberately do not protect everyone’s current position as such, because doing so would give the status quo exactly the veto we objected to at the start. Where it could be used CORE is a natural fit for situations that already call for unanimity, or for something close to it, and in which unanimity regularly ends in deadlock. NATO takes its deci- sions by consensus among its members, and the Council of the European Union still requires unanimity in areas such as foreign policy, defence and taxation. In both, a single member can hold up a decision that all the others support, and the fallback is that nothing happens. Citizens’ assemblies usually deliberate at length and then close with a vote, which hands the last word back to a majority; CORE could take the place of that vote. Siblings who inherit a house together and can sell it only if all of them agree can be stuck for years, and the legal fallback is often a court-ordered sale that serves none of them well. In each case CORE keeps unanimity as the goal and replaces the deadlock with a fallback that treats everyone alike. Two cautions apply. The first is that the uniform lottery is the right fallback only 5 where the default is meant to be neutral among the options. A criminal jury, which in several legal systems also has to be unanimous, shows where it is not. The presump- tion of innocence deliberately favors the defendant, and a fallback that convicted with a probability equal to the share of jurors favoring conviction would abandon it. A jury settling the amount of damages, by contrast, is a natural case. The second caution is that the members of such a body have to accept the commitment CORE requires, which includes being bound by a decision reached after they were removed. They also have to accept it in advance, for a whole class of decisions. A member who is free to decide case by case whether a question goes to CORE keeps the veto, since they will decline whenever the present arrangement suits them better than the lottery (Section 8.1). For the siblings, the commitment would have to come from the will or from the law of inheritance. For sovereign states it is a large step, and Section 13.3 discusses why such commitments are hard to obtain. Outline of the paper Sections 2 to 6 define the mechanism and prove the results about the endowment and the lottery. Section 7 separates what the mechanism guarantees from what it only uses as a benchmark, once smaller groups can reach agreements of their own, and compares what a single refusal can do in each variant. Section 8 discusses why people take part and what a decision can cover. Section 9 sets out what the impos- sibility theorems of Gibbard and Hylland establish for CORE, the narrower property the mechanism does have, how the gains from agreement are divided by bargaining, and why deliberation is left without formal rules. Section 10 asks who is entitled to take part in a decision. Section 11 offers an interpretation of the mechanism as a veil of uncertainty. Section 12 relates CORE to existing work, in particular to the non-deterministic consensus methods of Heitzig and Simmons, which use the same benchmark. Section 13 sketches how the mechanism extends to sortition and delega- tion, and Section 14 collects the problems that remain open. A note for readers of the earlier presentation. An earlier version of this work (Kōmyō and Claude 2025) defined each participant’s acceptance threshold by a recursion over subsets of participants. The recursion assumed that each subset reaches an agreement whenever one is available, but it did not say which agreement. Since that choice changes every value computed from it, the threshold was not well defined. Here we use the endowment as the benchmark, which requires no such choice. This does not make the underlying question disappear. In the sequential variant, what participants actually hold still depends on which agreements smaller groups would reach, and we leave that as an open equilibrium question (Section 14). What we offer instead are guarantees that hold whatever those agreements are. Three claims of the earlier work are withdrawn. Whether to accept a proposal is not a decision that can be made without regard to what others will do. The endowment tells a participant what they need not accept, and beyond that the decision belongs to bargaining (Section 9.4). An addendum on strategic manipulation (Kōmyō and Claude 2026) claimed that every participant can guarantee their endowment whatever the 6 others do, which fails in the sequential variant (Section 7.5). And the earlier work called the mechanism Rawlsian, a label we no longer use (Section 11.1). Readers who are new to CORE can pass over this note, since the paper does not otherwise refer to the earlier versions. 2. The mechanism Let 𝑉 be a finite set of 𝑛 participants. Each participant brings something into the decision, such as money into a fund, their own usage into a technical standard, or their labor into a project. We call everything brought in the pool , and we write 𝒜 for the set of outcomes that can be realized by acting on the pool. Section 8 discusses what this means. For now it is enough that 𝒜 contains only what the participants have at their disposal. Each participant 𝑖 evaluates outcomes by a von Neumann-Morgenstern utility func- tion 𝑢 𝑖 ∶ 𝒜 → ℝ , which extends to lotteries over 𝒜 by taking expectations. We write 𝑎 ∗ 𝑖 for an outcome that maximizes 𝑢 𝑖 over 𝒜 , call it 𝑖 ’s ideal , and write 𝑡 𝑖 = 𝑢 𝑖 (𝑎 ∗ 𝑖 ) for its value to 𝑖 . We assume that such an outcome exists, as it does when 𝒜 is finite, or compact with continuous utilities. If several outcomes are equally good for 𝑖 , we take 𝑎 ∗ 𝑖 to be the one 𝑖 would enact if left alone. Because ideals are taken within 𝒜 , they are always feasible: a participant’s ideal is the best that can be done with the pool, not the best they can imagine. Proposals may be lotteries. This matters for the substance of the mechanism, because the one proposal a group can always agree on is a lottery, as Section 5 shows. On a current set 𝑆 of participants, the procedure runs as follows. 1. Participants deliberate and may propose any outcome in 𝒜 or any lottery over 𝒜 2. If every member of 𝑆 accepts a proposal, it is enacted and the process ends. 3. Any member of 𝑆 may call for a removal. One member of 𝑆 is then drawn uniformly at random, possibly the caller, and removed. The process continues on the reduced set. 4. If only one participant remains, their ideal is enacted. Two features of this description are easily misread. The first is that step 3 is not triggered by a failure of step 2. Deliberation and the call are independent of each other, and both are available at any time: a participant may call for a removal while a promising proposal is under discussion, and a group may go on deliberating long after its first proposals have failed. The second is that the steps do not describe a vote. Nothing in CORE puts proposals to the group one at a time, to be accepted or rejected in turn. As soon as proposals are taken up in a fixed order, the order itself favors some of them over others. Deliberation is instead an open search that may arrive at an outcome everyone accepts, and it ends when the participants recognize that they have found one. It has no formal rules, because any such rule would set a constraint that could be exploited (Section 9.5). The language of proposals and acceptance serves only to model the result of deliberation. It does not prescribe how deliberation proceeds. 7 We call the procedure above the sequential mechanism. It has a one-shot variant, in which a call removes participants until only one remains, all in a single step, so that a stalled deliberation goes directly to the uniform lottery over the participants’ ideals. The one-shot variant is essentially the method that Heitzig and Simmons (2012) call “Full consensus or Random Ballot fall-back”. The two variants differ only in whether smaller groups get a chance to agree along the way. Section 7 examines what that difference costs and what it buys. Three further remarks clarify what kind of mechanism this is. First, invoking CORE is itself a commitment to reach a single outcome. One might ask why a group that disagrees does not simply split, with each faction carrying out its own plan. The answer is that splitting is not an alternative within CORE but the ungoverned situation CORE starts from. People can always come together or not, agree or not, and regroup, and doing so costs nothing. Splitting reveals which groups already exist, but it produces nothing beyond that. The difference shows up in what each participant falls back on. When splitting is the fallback, each participant mea- sures a compromise against what their own faction can do alone, and whoever does well alone has no reason to move. Under CORE the fallback is the lottery, which is worse for each participant than a good agreement, and it is this ordering that makes agreement worthwhile. Everything CORE adds to ungoverned interaction lies in the commitment to produce one outcome and adopt it. Second, a removal takes away a participant’s voice but not their stake. A removed participant takes no further part in deliberating or in accepting proposals, but what they brought into the pool stays there, and they receive whatever the remaining group enacts. This follows from the commitment just described, since a participant who could escape the outcome by being removed would never have been committed to it. It also has a cost, which Section 7 makes precise: what a removed participant brought can be disposed of by the others. Third, every agreement is an interim agreement. Reaching one settles the decision for now, not forever. Circumstances change, and an agreement may later need to be adjusted, which is done by a new process of the same kind, starting from the situa- tion the earlier agreement created. In that new process the earlier agreement is not protected as such, for the same reason that nobody’s current position is (Section 8.3). Within a single deliberation, too, the participants may agree on an interim result with- out ending the search: they may adopt a lottery that everyone prefers to the current fallback, and continue looking for something better. Such a step needs no rule of its own, since it is one more thing a group can agree on unanimously. From then on the adopted lottery is what the group has decided, and the continued search is a search for a unanimous amendment to it. Unless the agreement says otherwise, a call at this stage no longer removes anyone. It ends the search, and the adopted lottery is carried out. In the continued search every participant is therefore pivotal again, as in the one-shot variant (Section 7.3), which is the price of having locked in the gain. Section 9.4 shows why such steps matter for how bargaining converges. 8 3. The endowment Definition. For a participant 𝑖 in 𝑉 and a set 𝑆 of participants, the endowment of 𝑖 at 𝑆 is 𝑒 𝑖 (𝑆) = 1 |𝑆| ∑ 𝑗∈𝑆 𝑢 𝑖 (𝑎 ∗ 𝑗 ) It is the value to 𝑖 of the uniform lottery over the ideals of the members of 𝑆 . Equiv- alently, it is what 𝑖 receives under random dictatorship on 𝑆 , and what 𝑖 receives if removals continue from 𝑆 until one participant is left without any agreement being reached. The definition applies whether or not 𝑖 belongs to 𝑆 . This matters, because a removed participant still receives the outcome. Heitzig and Simmons (2012) use the same quantity as their benchmark. They call the uniform lottery over the voters’ favorites the benchmark lottery and define potential consensus options as the options everyone prefers to it. Three properties of the endowment are used throughout. It is selection-free: com- puting 𝑒 𝑖 (𝑆) requires only the participants’ ideals, with no assumption about which agreement any group would reach or about anyone’s strategy. It is the value of some- thing the procedure actually does: if removals run to the end without agreement, a single participant remains and their ideal is enacted, and since a sequence of uniform removals leaves each member of 𝑆 equally likely to be the survivor, 𝑒 𝑖 (𝑆) is exactly the value of this outcome to 𝑖 . And it is a benchmark rather than a guarantee. In the one-shot variant the endowment is exactly what each participant holds before an agreement is reached. In the sequential variant it is not, because smaller groups may reach agreements that change what everyone receives, whether present or removed. Section 7 shows that this difference can go in either direction and can be large, and it states what each participant is guaranteed in every case. 4. The call for random exclusion The call for a removal is the only move a participant can make alone. We first record a fact that holds exactly in both variants, and then a fact about the call measured against the benchmark. Observation (Caller independence). The effect of a call does not depend on who makes it. Every call removes a uniformly random member of 𝑆 , and the process continues from the reduced set without any record of who called. Simple as it is, this observation carries a lot of weight. No participant or group has an instrument that others lack. Whatever advantage a call gives anyone, it gives them equally whether they made the call themselves or their opponent did. Theorem 1 (Fair call). Let |𝑆| = 𝑘 ≥ 2 and consider a single call at 𝑆 . For every 𝑖 in 𝑉 , the expected endowment after the call equals 𝑒 𝑖 (𝑆) 9 Proof. Write 𝑇 = ∑ 𝑗∈𝑆 𝑢 𝑖 (𝑎 ∗ 𝑗 ) , so that 𝑒 𝑖 (𝑆) = 𝑇 /𝑘 With probability 1/𝑘 the removed participant is 𝑗 , which leaves 𝑆 ∖{𝑗} with 𝑒 𝑖 (𝑆 ∖{𝑗}) = (𝑇 −𝑢 𝑖 (𝑎 ∗ 𝑗 ))/(𝑘−1) This holds whether or not 𝑗 = 𝑖 , since a removed participant still receives the outcome. The expectation is 1 𝑘 ∑ 𝑗∈𝑆 𝑇 − 𝑢 𝑖 (𝑎 ∗ 𝑗 ) 𝑘 − 1 = 𝑘𝑇 − 𝑇 𝑘(𝑘 − 1) = 𝑇 𝑘 = 𝑒 𝑖 (𝑆). The argument uses no property of the utilities. □ A call is therefore a fair gamble in endowments. It spreads them out without moving their average, and it transfers nothing, to the caller or to anyone else. Since the util- ities already contain each participant’s attitude to risk, a gamble that is fair in utility is neither good nor bad for them: measured against the benchmark, a call leaves ev- eryone exactly where they were in expectation. Where risk does make a difference is in the comparison between the lottery and a compromise, which Section 6 takes up. In the one-shot variant the theorem is a statement about what participants actually hold. In the sequential variant it says only that calling does not move the benchmark, and Section 7 shows that a call can move what participants actually hold. Corollary 1.1 (Any number of calls). The expected endowment after any sequence of calls equals 𝑒 𝑖 (𝑆) , whoever makes the calls and however the decision to call again depends on who has been removed so far. In particular, removals that run to the end are worth 𝑒 𝑖 (𝑆) to every participant. Proof. By Theorem 1, the expected endowment after each further call equals the en- dowment before it, whatever has happened up to that point. Since at most 𝑘 − 1 calls can be made, the claim follows by induction on the number of calls. □ Corollary 1.2 (No gain against the benchmark). No participant can raise their endow- ment by calling. In the one-shot variant, a participant can therefore always refuse an offer worth less than their endowment. Calling instead costs them nothing in expec- tation and secures exactly their endowment, so that nobody has reason to accept less. It does not secure more. What a participant obtains above their endowment depends on what the others are willing to concede, and each of them holds the same threat (Section 9.4). Corollary 1.3 (The caller must stand in the draw). Suppose instead that callers were exempt, so that one of the other 𝑘 − 1 participants is removed. Write 𝐶 𝑖 (𝑆 ′ ) for the value to 𝑖 of continuing from a set 𝑆 ′ , whatever the participants actually do there. Exemption then changes the caller’s expected value by 1 𝑘 [ 1 𝑘 − 1 ∑ 𝑗≠𝑖 𝐶 𝑖 (𝑆 ∖ {𝑗}) − 𝐶 𝑖 (𝑆 ∖ {𝑖})] that is, by the difference between the caller’s average value when surviving and their value when removed, divided by 𝑘 . Measured in endowments, where 𝐶 𝑖 = 𝑒 𝑖 , this 10 equals 𝑡 𝑖 − 𝑒 𝑖 (𝑆) (𝑘 − 1) 2 which is never negative, and positive for anyone who is not indifferent among the ideals present. Proof. With the caller standing in the draw, the caller’s expected value is (1/𝑘) ∑ 𝑗∈𝑆 𝐶 𝑖 (𝑆 ∖ {𝑗}) . With the caller exempt, it is (1/(𝑘 − 1)) ∑ 𝑗≠𝑖 𝐶 𝑖 (𝑆 ∖ {𝑗}) Subtracting the first from the second gives the stated expression. For en- dowments, (1/(𝑘 − 1)) ∑ 𝑗≠𝑖 𝑒 𝑖 (𝑆 ∖ {𝑗}) = ((𝑘 − 2)𝑇 + 𝑡 𝑖 )/(𝑘 − 1) 2 and 𝑒 𝑖 (𝑆 ∖ {𝑖}) = (𝑇 − 𝑡 𝑖 )/(𝑘 − 1) , with 𝑇 as in Theorem 1. Substituting gives (𝑘𝑡 𝑖 − 𝑇 )/(𝑘(𝑘 − 1) 2 ) = (𝑡 𝑖 − 𝑒 𝑖 (𝑆))/(𝑘 − 1) 2 , which is nonnegative because 𝑡 𝑖 is at least every 𝑢 𝑖 (𝑎 ∗ 𝑗 ) □ Measured against the benchmark, then, exempting callers would make calling prof- itable for almost everyone, and most of all for those whose ideal lies furthest from the rest of the group, who would call constantly. Exemption would also turn the call into a race. A participant who cannot be removed by their own calls is the more likely to be left at the end the more often they call, and one who called before anyone else each time would be certain of it. Requiring callers to stand in their own draw is what makes the only individual move neutral with respect to who makes it. In the sequential vari- ant the general expression is typically positive as well, but this is not guaranteed. It can be negative when the others would agree on something the participant likes once the participant is gone, but would not accept it while the participant is present, be- cause the participant’s ideal gives them a gamble that they prefer and the participant fears. Such cases are special. In the ordinary case, the compromise the others would reach without a participant is also on offer while that participant is present, and the participant can accept it or ask for more. 5. Consensus and the fallback Theorem 2 (The lottery as a proposal). Let 𝐿(𝑆) be the uniform lottery over {𝑎 ∗ 𝑗 ∶ 𝑗 ∈ 𝑆} . Then 𝑢 𝑖 (𝐿(𝑆)) = 𝑒 𝑖 (𝑆) for every 𝑖 Proof. This follows directly from the definition of the endowment and the expected- utility extension. □ In the one-shot variant, 𝐿(𝑆) is exactly what a stalled deliberation produces. There is therefore always a proposal that everyone weakly accepts, and adopting 𝐿(𝑆) by agreement is at least as good for everyone as letting the deliberation stall, since both lead to the same distribution over outcomes. The fallback does its work counterfactu- ally: it fixes what any agreement has to beat, without needing to occur. Heitzig and Simmons (2012) prove the matching equilibrium result for their method. Using a strong equilibrium concept in which coalitions can coordinate their devia- tions, they show that whenever some option is preferred by everyone to the bench- 11 mark lottery, the equilibria are exactly the efficient ways of electing such options by full consensus, so that no equilibrium outcome is worse for anyone than the bench- mark. Their result concerns a fixed set of options and one-shot ballots rather than open deliberation, but its structure is that of our one-shot variant. It supplies some- thing we do not prove independently here: that agreement is not merely available but an equilibrium, and that there is no equilibrium of any other kind. Whether partici- pants arrive at it is a further question, since with several such options they still have to settle on one (Section 9.4). In the sequential variant, an analogous proposal exists but is no longer selection-free. Consider the lottery over outcomes that a call would actually lead to, given whatever agreements smaller groups would reach. It gives each participant exactly what they would hold after a call, so everyone weakly accepts it too. But what that lottery is depends on which agreements smaller groups would reach, which is precisely the question the endowment was introduced to avoid. And 𝐿(𝑆) itself is no longer guar- anteed to be acceptable, since participants who expect to do well in smaller groups will reject it (Section 7). 6. Where the gains from agreement lie Theorem 3 (Support). The fallback lottery gives each outcome a probability equal to the share of participants whose ideal it is. In particular, an outcome that is nobody’s ideal has no chance under the fallback and can be reached only by agreement. Proof. The fallback enacts the ideal of the last remaining participant, and each partic- ipant is equally likely to be the last. □ This tells us where the value of deliberation lies. The lottery counts first choices and nothing else. How much the other participants value an outcome plays no role in the weight it receives, so whatever a group gains by agreeing rather than gambling, it gains by giving some outcome more weight than its share of first choices. This may be a compromise that nobody ranks first and that the lottery never selects, or one participant’s ideal that the others also value highly. In both cases the gain is invisible to the lottery and can be found only through the search that deliberation performs. An example of Dutta, Peters and Sen (2007) shows how much is at stake. A thousand participants each have a different favorite among a thousand and one options, and every one of them ranks the remaining option second. Suppose each values their own favorite at 1, the common second choice at 0.9, and everyone else’s favorite at 0. The lottery over ideals is worth 0.001 to each participant, and it never selects the option that everyone ranks second, while agreement on that option is worth 0.9 to each. The example also shows why unanimity matters for reaching such an option. Under the COWPEA Lottery of Pereira (2023), a random ballot is drawn and only the candi- dates it approves are kept, further random ballots remove the candidates they do not approve, and a ballot that would remove all remaining candidates is ignored. If every participant approves their favorite and the common second choice, the second choice 12 is elected for certain. But a participant who approves only their favorite still gets the second choice whenever another ballot is drawn first, and gets their favorite in the one case in a thousand in which their own ballot comes first. That is worth 0.9001 rather than 0.9, so every participant gains by withholding approval of the compro- mise, and if all of them do, the result is the lottery again. The compromise is a public good on which each participant can free-ride. In the one-shot variant of CORE this cannot happen. The compromise is enacted only if every participant accepts it, so refusing it does not let anyone keep the compromise and add a chance of their favorite. Without their acceptance there is no compromise, and what remains is the lottery, worth 0.001. Unanimity makes every participant pivotal, and a pivotal participant has nothing to free-ride on. A participant who wants more has to ask the others for it, which turns free-riding into open bargaining (Section 9.4). Heitzig, Simmons and Constantino (2025) identify the same free-rider problem and address it in MaxParC with conditional commitments to approve compromise options. Being pivotal has a price, however, since a participant without whom there is no compromise can also hold it up. In the sequential variant both effects are weaker. Section 7.3 compares the variants in this respect. Theorem 3 also explains why CORE is not an aggregation rule in the classical sense. An aggregation rule maps a profile of preferences over a fixed set of alternatives to an element of that set. In CORE the gains may lie in alternatives that were on no list at the outset, and generating them is part of what the procedure does. Impossibility theorems about aggregation over a fixed set of alternatives therefore do not describe CORE. This is not a way of evading those theorems but a description of what kind of object CORE is, and Section 9 discusses the results that do apply. The advantage of agreement over the lottery is conditional, and the condition is worth stating plainly. Proposition 1 (Zero-sum reduction). If no lottery over 𝒜 is weakly preferred to 𝐿(𝑆) by every member of 𝑆 and strictly preferred by at least one, then every lottery that all members of 𝑆 weakly prefer to 𝐿(𝑆) gives each of them exactly their endowment, and agreement can do no better than the fallback. This is the case, for instance, when two participants face two outcomes on which their preferences are exactly opposed, with utilities normalized to 0 and 1. Every lottery then gives the two participants utilities that sum to 1, and CORE reduces to random dictatorship. Proof. The first claim restates the assumption. In the example, a lottery that places probability 𝑝 on the first participant’s preferred outcome gives the two participants utilities 𝑝 and 1 − 𝑝 , so no lottery makes both better off than (1/2, 1/2) □ The proposition marks the scope of what CORE can do. The advantage of consensus over randomness extends exactly as far as compromise is possible. Where the conflict is purely about how to divide something fixed, deliberation has nothing to find and the mechanism returns the fair lottery. Where compromise exists, deliberation is the only way to reach it. 13 At the other end stands a large class of decisions in which a gain is certain to exist. Observation (Averaging beats the lottery). Suppose 𝒜 is a convex set, so that out- comes can be averaged, as with the divisions of a budget or the positions on a scale, and suppose every 𝑢 𝑖 is concave on it. Let ̄ 𝑎(𝑆) = 1 |𝑆| ∑ 𝑗∈𝑆 𝑎 ∗ 𝑗 be the average of the ideals. Then 𝑢 𝑖 ( ̄𝑎(𝑆)) ≥ 𝑒 𝑖 (𝑆) for every 𝑖 The inequality is strict for every participant whose utility is strictly concave, unless all the ideals coincide. Proof. By Jensen’s inequality, the value of an ave