| paper | d4-conf-aaai-2026-1836 |
| authors | — |
| venue | ? |
| filed under | unclassified |
| judged by | gpt-6-luna / xhigh (triple__luna__xhigh__c6r1) |
| judge confidence | medium |
| authors would recognise it | yes |
other→P discrete result → its continuous mirror
Answer: Universal EVS convergence is false. On inputs admitting the specified EVS limit, evaluation is polynomial-time; the limit is admissible and has the policy-determined minimum objective \(J=0\). · discrete side: other · adjudication: proved_with_minor_gaps
full proof & adversarial review →
Theorem 4
statement extracted from the paper’s text layer
Given a finite set of complete contribution roles \(T\), rational role masses \(\mu\), and a succinct monotone Boolean value functional \(V:[0,1]^T\to\{0,1\}\) on arrived role masses, define finite replicas by \(v_n(S)=V((|S\cap N_t|/n)_{t\in T})\), with \(|N_t|/n\to\mu_t\). Under uniform arrival orders and tagged-agent delay deviations, does EVS converge to a measurable per-capita policy that minimizes the limit of \(n\,\mathbb{E}_{\pi}[\sum_i(\phi_i(v_n,\pi)-\mathrm{SV}_i(v_n))^2]\) among anonymous policies satisfying the limiting SF, OIR, I4EA, and MOS conditions, and can that policy be computed from the chosen encoding of \(V\)?
A large project workforce represented by masses \(\mu_t\) of complete contribution roles, with project value \(V\) determined by arrived role masses. Finite replicas supply per-capita shares and Shapley benchmarks; tagged-agent deviations retain the paper's early-arrival incentive question.
The paper gives no succinct representation for the value functional on role masses, so the continuous problem's computational complexity depends on a further modeling choice; tagged-agent delay incentives also need a convergence argument.
fatal: False
The mirror covers Theorem 4's expected Shapley-distance optimality and Theorem 3's egalitarian-welfare optimality; it leaves the general WVS guarantee of Theorem 2 and the paper's treatment of unsolvable games without a limit classification.
The strongest positive lead is the paper’s optimization of in-order fairness, but there is an important limit: it has no named computational-complexity result of the kind your anchor rule requires. Its theorems characterize mechanisms and their fairness guarantees; they do not say that an underlying problem is in P, hard, or otherwise classify its complexity.
The closest result is Theorem 4, proved here: for any solvable \(0\)-\(1\) monotone game, EVS has no greater expected squared Shapley distance than any anonymous mechanism satisfying SF, OIR, I4EA, and MOS. Theorem 3, also proved here, says that for each solvable game and joining order, EVS’s egalitarian welfare is at least that of any mechanism satisfying SF, OIR, I4EA, and MOS. And Theorem 2, proved here, establishes those four properties for WVS when its weight function is weakly decreasing. These are meaningful mechanism-design results, but none supplies the required complexity anchor or a runtime bound.
There is a plausible population regime behind Theorems 3 and 4: a large project workforce made up of repeated contribution roles, where many contributors have identical preferences, costs, and contribution behavior. A type would be a complete such role; \(\mu_t\) would be its fraction of the workforce. A natural continuous question is whether an online policy can allocate a project’s \(0\)-to-\(1\) value across the arrived critical mass, satisfy continuum versions of the paper’s fairness and incentive conditions, and minimize the limit of the scaled expected squared deviation from Shapley shares in finite replications of that workforce. EVS suggests a candidate solution.
That is a real population mirror of the paper’s central fairness objective, rather than a continuization of arrival time alone. But making it precise requires defining the replicated games and showing their Shapley benchmarks and EVS allocations converge. In a continuum, an individual arrival has zero mass, so the paper’s pivotal-player and delay arguments do not transfer automatically. I would expect the first question to be whether the EVS optimality result survives that limit; the paper does not establish a tractability or hardness classification for it.
So this is the best conceptual case for a continuous mirror, and Theorem 4 is its strongest named motivation. Under the requested anchor criterion, though, it is not an admissible anchor: I can’t honestly turn it into a complexity result the authors proved.
The strongest objection is that none of the proponent’s three anchors is a computational result. Theorem 2 proves that WVS satisfies fairness and incentive properties; Theorems 3 and 4 prove optimality for two fairness objectives. The paper gives no encoded computational problem, complexity classification, runtime bound, or algorithmic guarantee. Algorithm 2 describes a mechanism, but the game \(v\) is an arbitrary monotone function on \(2^N\), with no representation or cost model for finding its critical players or minimal critical prefix. There is therefore no complexity claim here to carry over to continuous populations.
The best population scenario is repeated contribution roles in a large project. One could describe the workforce by type masses, define a value function \(V(x)\) over coalition masses, and ask for a limit of per-capita EVS shares and scaled Shapley-distance welfare under random arrivals. That is a coherent candidate model, and it takes Theorem 4 seriously. But it adds the ingredients the paper does not supply: a representation for \(V\), a law for type arrivals, a normalization for individual shares, and a limit definition for the Shapley benchmark. Those choices define a new problem. Depending on the chosen representation, computing the limit could be simple or hard; the paper gives no reason to privilege one computational question over another.
The same gap affects Theorem 3. In an atomless population, “the minimum share among critical players” needs a replacement such as the essential infimum of per-capita allocation rates over critical mass. That may be a sensible continuous fairness objective, but it is a new objective on a mass-valued game. To make it computational, one still has to choose how the mass game is encoded and what counts as an input.
Theorem 2’s incentive condition has a related modeling cost. I4EA compares what a particular player receives when that player changes position while everyone else’s order stays fixed. A measure over types and arrival times does not record that single-player deviation: moving one atomless agent changes no mass. A tagged-agent or finite-replica formulation could preserve the comparison, but then the continuous state must be supplemented with individual-level deviation data and a per-capita payoff convention. That might support a useful limit study; it does not reveal a computational result in this paper that the programme can mirror.
So I would reject this paper as a strong *computational-complexity mirror* candidate: its named results concern axioms and mechanism optimality, and making any of them computational requires substantial new modeling choices. The universal claim that no scenario could yield worthwhile continuous work is weaker. Repeated-role threshold projects are plausible, and a careful limit analysis could be worthwhile. The honest negative case is that this would be a new continuous mechanism-design project, not a continuization of a computational result established by this paper.
The adversarial triple: the proponent anchors on up to three named results; the opponent sees that case and must defeat every anchor; the judge decides which case convinced it. These are the pipeline’s own outputs, generated by tools/triple_run.py — no human edited them. The paper’s own text is not reproduced here beyond the quoted statement above.