| paper | d4-conf-aaai-2021-1783 |
| authors | — |
| venue | ? |
| filed under | unclassified |
| judged by | gpt-6-luna / xhigh (triple__luna__xhigh__c6r1) |
| judge confidence | high |
| authors would recognise it | yes |
The paper defines AUT and reports its predictive accuracy, but contains no numbered theorem, lemma, corollary, or proposition asserting a computational result. Aggregating AUT predictions over type masses is a plausible descriptive forecast, but it cannot supply the missing computational anchor.
fails bit a — no named computational result to mirror
No named computational result is covered; a type-level aggregation would describe AUT forecasts, while the paper's behavioral comparisons remain empirical.
The best population-level foothold is AUT’s prediction of how people cast approval ballots. In a large online committee election with a small set of recurring utility profiles, let a voter type \(t\) be a complete utility vector and, if relevant, that type’s AUT parameters. Let \(\mu_t\) be the fraction of voters of type \(t\). Given the shared election context—current approval scores, number of winners, and missing-ballot assumptions—AUT predicts a ballot for each type; aggregating those predictions gives the society’s ballot distribution. That is a plausible continuous version of the paper’s stated interest in predicting behavior at both individual and group level.
But the paper has no numbered Theorem, Lemma, Corollary, or Proposition asserting a computational result. Its finding that AUT predicts ballots more accurately than the tested alternatives—including the reported accuracies of \(58.1\%\), \(68.0\%\), and \(67.9\%\) for one-, two-, and three-winner settings—is an empirical result, not a complexity anchor. The complexity claims in Related Work are attributed to other papers, and cannot serve as this paper’s anchors.
So there are no qualifying anchors for which to state the required continuous prover questions. Turning the population-level AUT forecast into a problem of fitting parameters to aggregate ballot data would be a possible new research question, but it would not mirror a named computational result of this paper. The continuous-population setting is plausible; the paper does not supply the computational result needed to make the requested anchor-based case.
The AUT result is the paper’s only plausible foothold. A sensible population version would put voters with the same utility vector and AUT parameters into a type \(t\), assign each type mass \(\mu_t\), and predict the mass casting ballot \(b\) as
\[
q(b)=\sum_t \mu_t\,\mathbf{1}\!\left[M^{AUT}(t)=b\right].
\]
That fits the authors’ interest in predicting behavior at group level, so I would not call the model incoherent. But the paper’s reported prediction accuracies are empirical findings, not computational results: it states no theorem or algorithmic guarantee about predicting ballots, fitting AUT, or computing the aggregate forecast. The complexity results in Related Work belong to other papers.
The stronger version of the proposed mirror still gives little for this programme to chart. AUT evaluates a score for each candidate and applies a threshold; for a given type and election context, it does not search over ballots or couple one type’s decision to another’s. Computing \(q\) is weighted aggregation of those predictions. With a fixed type support, increasing the population size mainly changes the weights. A complexity study of fitting parameters to a population would be a new learning question, while optimizing a campaign against the population’s response would add an objective and intervention absent from this paper. Neither mirrors a result the paper establishes.
There is also a limit to the strategic interpretation: in an atomless society, one voter has zero mass and cannot change the outcome. Giving a whole type’s mass the power to switch ballots would model a coordinated bloc, while the experiment studies individuals responding to a given tally. That weakens the case for a continuous version of the paper’s strategic setting, though it does not rule out useful descriptive population forecasts. So the negative case is strongest on scope: a continuous forecast is plausible, but this paper offers no computational result for it to mirror, and the obvious aggregate computation does not supply a substantial complexity question. I would not claim that no useful continuous model could ever be built; that universal claim is weaker than the evidence supports.
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.