| paper | d4-conf-ijcai-2020-0041 |
| authors | — |
| venue | ? |
| filed under | unclassified |
| judged by | gpt-6-luna / xhigh (triple__luna__xhigh__c6r1) |
| judge confidence | high |
| authors would recognise it | no |
Theorem 1, Theorem 2, and the corollaries establish incentive guarantees, not algorithms, hardness classifications, or computational formulations used to compute. The proposed minimum-audited-mass task adds an aggregate audit-allocation objective that the paper does not study, so it is not a computational mirror of these results.
fails bit a — no named computational result to mirror
The proposed audited-mass optimization is not established by Theorem 1, Theorem 2, or Corollaries 1 and 2; those results compare incentive guarantees and spot-check probabilities.
I cannot make a compliant anchor-based case from this paper: it has no named result classifying a computational problem as polynomial-time, NP-hard, FPT, or similar. Its numbered results concern incentive guarantees and equilibria, not computational complexity. Calling them complexity anchors would overstate what the paper proves.
The closest results are Theorem 1, proved in this paper, which says that for any universal mechanism and any positive extra effort cost, some multi-signal environment makes the truthful equilibrium fail to be Pareto dominant; and Theorem 2, also proved here, which gives a sufficient condition under which the peer-insensitive mechanism’s minimum spot-check probability \(p_{\mathrm{ds}}\) is no greater than the minimum \(p_{\mathrm{Pareto}}\) for a qualifying spot-checking peer-prediction mechanism, for every \(c_E\ge 0\). Corollaries 1 and 2 apply that comparison to listed families of peer-prediction mechanisms. These are meaningful incentive results, but none asserts a complexity classification.
A plausible continuous-population follow-up would be Minimum Audited Mass for Truthful Peer Evaluation. An instance would give a finite set \(T\) of evaluator types and rational masses \(\mu_t\) summing to one. Each type specifies its task assignments, effort cost, and finite signal distributions; a trusted evaluator’s signal distribution and a peer-insensitive payment rule are also given. A fraction \(p\) of evaluation mass is spot-checked. The question is to find the smallest \(p\in[0,1]\) for which acquiring the high-quality signal and reporting it truthfully is a dominant strategy for every type. The solution is \(p\) together with a certificate of the required incentive inequalities.
A credible regime is a large peer-grading or review platform with many evaluators but relatively few evaluator profiles: for example, cohorts sharing a task, effort cost, and signal quality. Then the number of agents \(n\) can be much larger than the support size \(\tau=|T|\), and \(\mu_t\) is the fraction of the evaluator population in each profile. Types describe distributions and costs, not identical realized signals.
For explicit finite signal and payment tables, the dominant-strategy version looks likely to reduce to finitely many payoff inequalities. The paper does not establish that algorithmic result, though, and its equilibrium-comparison theorems do not determine the complexity of a continuous-population version. The main weakness is that the paper’s contribution is about strategic guarantees; moving to population masses may be a sensible high-multiplicity regime, but it does not reveal a computational landscape of the kind ChoCo is meant to chart.
The strongest negative case is at the programme’s entry gate: this paper gives no named computational problem whose complexity or algorithms a continuous mirror could chart. Theorem 1 is an impossibility result about equilibrium incentives. Theorem 2 and its corollaries compare incentive guarantees and spot-check probabilities. None defines a computational input and asks for an algorithm, or classifies such a problem as tractable or hard. Those are important results, but making their settings continuous would continue the game-theoretic analysis rather than supply the computational question ChoCo is looking for.
Theorem 1 has a plausible large-population setting: a review platform or course with many evaluators sharing task assignments, effort costs, and signal distributions. That makes repeated types sensible. But the impossibility turns on what an individual can learn, report, and coordinate on. Replicating evaluator types changes their population shares; it does not create a computational task. A continuum version could restate the result for masses of evaluator types, but its central content would still be the existence of an uninformative equilibrium under the paper’s signal assumptions.
Theorem 2, Corollaries 1 and 2, and the proposed “Minimum Audited Mass” mirror have a more specific problem. The paper’s \(p_{\mathrm{ds}}\) and \(p_{\mathrm{Pareto}}\) are spot-check probabilities in agents’ payoff comparisons. Under a common audit probability, type mass does not enter those comparisons. With finite signal and payment tables, the truthful-incentive conditions reduce to finitely many inequalities: for each type and deviation, the expected gain from a trusted report must cover the effort cost. The minimum \(p\) is obtained by satisfying those inequalities. If audit rates vary by type and the objective is total audited mass, the objective becomes a mass-weighted sum of per-type rates; the constraints remain typewise. That is a possible optimization model, but it adds an aggregate audit-allocation objective rather than continuizing the paper’s comparison theorem.
A stronger extension could choose type-specific audit rates and payments jointly, perhaps with a shared audit budget or report-sensitive checks. Population masses could then affect the principal’s cost. Yet the hard question would be the new mechanism-design problem created by those choices. It would not arise from a computational result in this paper, and the authors’ stated comparison of peer-insensitive and peer-prediction mechanisms would not identify the computational landscape of that extension.
The negative case has a real limit: large peer-grading and review populations can plausibly have many agents per evaluator type, so I cannot honestly argue that no sensible continuous population exists. I can argue that these anchors provide little reason for ChoCo to spend time on a continuous mirror of this paper. Any worthwhile computational project appears to require adding an optimization objective and mechanism-design choices beyond the paper’s proved results.
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.