| paper | d4-conf-atal-2018-0217 |
| authors | — |
| venue | ? |
| filed under | unclassified |
| judged by | gpt-6-luna / xhigh (triple__luna__xhigh__c6r1) |
| judge confidence | high |
| authors would recognise it | no |
Theorem 4.6 is a named computational result under the stated gate: it guarantees budget feasibility for the one-shot heuristic. But the paper’s elicitation objective and matrix completion concern one user, and the proponent supplies no concrete type-mass question anchored to that result. Pooling costs or reconstruction error creates a new population allocation problem, so no continuous analogue that the authors would recognize survives.
fails bit b-changes — the continuum limit changes the question
The paper’s low-rank structure captures similarities across tasks in one user’s matrix, not similarities across users; making population mass matter requires a new pooled, cross-user allocation model.
fatal: True
Theorem 4.5 and Theorem 4.6 cover bother-budget feasibility for one-shot question selection, but give no elicitation-accuracy guarantee or complexity classification; the matrix-completion experiments remain outside the proposed mirror.
The strongest positive intuition is a large smart-home fleet managed by one DSM service: each home has the same device and time-slot menu, and the service wants to learn enough about preferences to produce useful schedules while keeping average bother within budget. Its population could be large relative to the number of recurring preference profiles. That makes population-level elicitation a plausible research direction.
But I cannot tie that case to a qualifying anchor in this paper. The closest named result is Theorem 4.6, proved here: if the one-shot heuristic returns \(Q^*\) for inputs \(Q,M,B,w\), then the bother cost of asking \(Q^*\) is at most \(B\). Theorem 4.5, also proved here, gives a sufficient condition on the discounted bother-cost sum for the total cost to stay within budget. Neither result establishes that a problem is in P, hard, or otherwise classifies its complexity. The paper’s 0-1 knapsack formulation and matrix-completion optimization are used by the system, but the paper does not give a named complexity result for the underlying preference-elicitation problem. Its main comparisons are empirical.
The high-multiplicity mirror also has a serious modeling gap. If a type includes a home’s complete preferences, as the programme’s definition requires, then homes of the same type have the same answers; learning a preference from one home could inform the others. If homes differ enough that their answers must be elicited separately, they are different types. The paper’s objective is to reconstruct one user’s preference matrix, so replacing it with an average over a fleet may change the problem in a way the authors would not accept.
So the fleet scenario is a lead for a possible extension, but there is no honest anchor-to-mirror pair here, and I cannot assign it a complexity class as a mirror of a result in this paper.
The proponent does not actually establish an anchor for a continuous mirror. Theorem 4.5 gives a sufficient budget check, and Theorem 4.6 says the one-shot heuristic’s selected questions stay within budget. Neither classifies the complexity of preference elicitation or gives an algorithmic guarantee about the quality of the elicited preferences. The paper’s knapsack formulation and matrix completion are computational components, but its main comparison is empirical. So the strongest negative case is that there is no named complexity result here for the programme to mirror.
The best version of the fleet idea would define each user’s complete preference matrix, bother parameters, and question responses as their type, then ask how to minimize population-average reconstruction error under a population-level bother budget. But the paper’s matrix completion is within one user’s matrix: its low-rank assumption concerns similarities between that user’s tasks, not similarities across users. If a type’s members really are indistinguishable, one representative’s answers can be reused, so increasing the type’s mass changes its weight in the aggregate objective but does not create more elicitation work. If answers cannot be reused, then the model needs an additional source of individual variation—such as noisy answers, privacy limits, or user-specific interaction constraints—that the paper does not study.
One can lift the budget guarantee to a population by applying it separately to each type and averaging the costs; that is essentially the existing result repeated across types. A shared budget that lets the system allocate questions unevenly across types is more interesting, but it changes the task into population-level resource allocation. That could be a worthwhile extension, yet it is not a continuous mirror of Theorems 4.5 or 4.6, and the paper does not establish a complexity result for that extension.
This is a strong case against *anchoring a mirror in this paper’s results*, but it does not prove the universal claim. A large service with recurring, genuinely shared preference profiles is plausible, and weighted aggregate elicitation could be worth studying. The case is weakest exactly there: it rules out a direct, result-based mirror more convincingly than it rules out a worthwhile new population-level problem.
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.