| paper | d4-conf-atal-2020-0355 |
| authors | — |
| venue | ? |
| filed under | unclassified |
| judged by | gpt-6-luna / xhigh (triple__luna__xhigh__c6r1) |
| judge confidence | high |
| authors would recognise it | yes |
The proposed type-mass MLE is a coherent direct population version, and its weighted-sign rule follows from the paper’s formula. But Case 3 is presented as a case derivation, not as a qualifying numbered result, and the empirical comparisons state no computational guarantee. The paper therefore has no qualifying anchor for a continuous computational mirror.
fails bit a — no named computational result to mirror
The candidate mirror covers only the two-alternative ordinal MLE in Case 3; it leaves the many-alternative approximations, learned rule, and experiments aside.
There is no admissible formal anchor in the supplied paper: it contains no numbered Theorem, Lemma, Corollary, or Proposition establishing a complexity result. It does derive aggregation rules and report experiments, but it proves no result of the form “this problem is in P,” “NP-hard,” or similar. So I can’t honestly give a mirror *per qualifying anchor*.
The closest substantive result is Case 3, though it does not meet that formal requirement. The paper considers two alternatives, ordinal votes, independent Gaussian observations, and voter-specific arm counts. It derives an MLE winner rule: voter \(i\)’s vote is weighted by
\[
s_i^{-1}=\sqrt{\frac{c_{i1}c_{i2}}{\sigma_2^2c_{i1}+\sigma_1^2c_{i2}}},
\]
and the sign of the weighted vote determines which alternative the MLE favors. This is an analytic aggregation result, not a complexity classification.
If Case 3 were accepted as an anchor, a plausible continuous mirror would be: there are two alternatives and a finite set \(S\) of observed voter types \(t=(y_t,s_t)\), where \(y_t\) records the ranking and \(s_t\) is its known noise scale. The instance gives rational masses \(\mu_t\) summing to one. For a truth difference \(\Delta\), define
\[
\ell_\mu(\Delta)=\sum_{t\in S}\mu_t\left[y_t\log\Phi(-\Delta/s_t)+(1-y_t)\log\Phi(\Delta/s_t)\right].
\]
The task is to return the MLE winner, including a tie if the optimum has equal truth values. Equivalently, following Case 3’s rule, the winner is determined by the sign of \(\sum_t\mu_t(1-2y_t)/s_t\).
A credible regime is a large expert panel repeatedly evaluating the same two proposals, with members assigned a small number of standardized evidence-exposure profiles. The complete type includes the observed ranking and exposure profile; many panelists share each type, so the number of panelists \(N\) can greatly exceed the support size \(\tau\). The continuous society is the distribution of those types. Authors would likely recognize this as a population version of their two-alternative MLE aggregation problem.
I would expect this particular mirror to be tractable: its winner rule reduces to a weighted sum over the \(\tau\) types. It covers only the two-alternative ordinal case, not the paper’s many-alternative approximations, learned neural rule, or empirical comparisons. A natural next question is whether the exact many-alternative ordinal MLE—without the paper’s naive-independence approximation—admits a useful continuous optimization or complexity characterization.
The weak point is that this is a statistical aggregation problem, and the paper does not frame it as a computational-complexity question. The proposed continuum mainly replaces a finite sum over voters with a weighted sum over type masses; the authors may see that as routine population likelihood analysis rather than a new complexity problem. That makes it a plausible, narrow mirror, but not a strong programme anchor.
The strongest case against is that the proponent supplied no qualifying anchor. The paper has no named theorem, lemma, or corollary classifying the complexity of a computational problem. Case 3 derives an MLE aggregation rule; the other cases derive or approximate rules and report empirical performance. So Case 3 is not a complexity result to mirror under the stated criteria.
That objection does not make its continuous version incoherent. A large panel could contain repeated groups with the same exposure profile and the same observed ranking. Taking type \(t\) to be that profile and ranking, the society’s masses define a weighted likelihood, and the authors would plausibly recognize the resulting MLE as their two-alternative problem. This defeats objections based on lack of multiplicity or on the types forgetting relevant information. The negative case can say only that this mirrors an analytic rule, not a computational result; it cannot honestly show that the population question is worthless.
The suggested exact many-alternative MLE is a stronger candidate, though it is not a mirror of a proved result. For types \(t=(\pi,c)\), with ranking \(\pi\) and exposure counts \(c\), one can ask to maximize \(\sum_t \mu_t \log P_t(\pi\mid V)\) over the latent arm means \(V\), where \(P_t\) is the probability of that ranking under the paper’s Gaussian observation model. Repeated standardized exposure groups make a plausible high-multiplicity setting. The paper itself leaves the exact many-alternative case unresolved and uses an approximation, so I cannot defeat this candidate on the grounds the question is already answered or that the continuous population is nonsensical.
So the proponent’s formal anchor fails, but their closing question survives as a potentially worthwhile new computational problem. The honest negative case is therefore weak: the paper offers no qualifying result to mirror, but the material does not support the universal claim that no worthwhile continuous mirror exists.
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.