| paper | d4-conf-aaai-2024-2625 |
| authors | — |
| venue | ? |
| filed under | unclassified |
| judged by | gpt-6-luna / xhigh (triple__luna__xhigh__c6r1) |
| judge confidence | high |
| authors would recognise it | no |
Neither cited theorem asserts an algorithm, a complexity bound, hardness, or a computational formulation used to compute; each proves a property of an aggregation functional. The proposed subset detector is a new computational question over the paper's existing weighted theory support, not a continuous-population analogue of a named computational result.
fails bit a — no named computational result to mirror
The paper's fanaticism classifications for MEC, Maximin, k-trimmed Highest Mean, and Highest Median have no continuous computational counterpart anchored in a named computational result.
The strongest positive case is a narrow one: the paper has a natural population interpretation and a clear computational question one could build from it. But it contains no named computational result of the required kind—no theorem or proposition classifying a problem as NP-hard, polynomial-time solvable, FPT, or similar. So there is no valid anchor for a continuous mirror under your rules.
The nearest result is Theorem 12, proved in the paper: MEC is Pascalian. In full, for every \(k\in(0,0.5)\), every ethical framework and every action set, MEC can be extended with added theories whose total credence is at most \(k\), so that those theories form a dominant subset: the extended framework gives the same action ordering as the added theories alone, and a different ordering from the original theories alone. Theorem 15, also proved here, gives the contrasting result that Highest Median is not \(k\)-fanatical for any \(k\in(0,0.5)\). These are results about aggregation behavior, not computational complexity.
A concrete computational question suggested by Theorem 12 is Minimum Dominant-Mass Detection under MEC. An instance has a finite action set \(A\), a support \(T\) of complete ethical-evaluation types, each with a rational vector \(u_t:A\to\mathbb Q\), and rational masses \(\mu_t>0\) summing to \(1\). Find a nonempty proper subset \(S\subset T\) of minimum mass \(\sum_{t\in S}\mu_t\) such that MEC’s ordering on the whole society equals its ordering on the normalized restriction to \(S\), but differs from its ordering on the normalized complement \(T\setminus S\). A threshold version asks whether such a subset exists with mass at most a given \(k<0.5\).
A plausible regime is a large AI deployment with many stakeholders but relatively few distinct, complete evaluation profiles: mass is the fraction of stakeholders of each type, with the number of people much larger than the support size \(\tau=|T|\). Theorem 12 motivates asking whether a small group determines the outcome; the authors’ formal notion of a dominant subset makes this question recognizable as an extension of their problem. For this proposed optimization problem, exhaustive subset enumeration gives an \(2^\tau\operatorname{poly}(|A|,L)\) algorithm, hence FPT in \(\tau\). NP-hardness and a Class C classification are plausible because the search is over subsets of types, but I would not claim either without a reduction.
The weak point is decisive for the requested grading: the paper already represents ethical theories as weighted “individuals,” with credences summing to one, and describes those credences in terms of societal advocacy. That is already close to a distribution over types. The proposed detection problem is a computational follow-on to its definitions, not a continuous counterpart anchored in a named complexity result.
The strongest objection is categorical: this paper has no named result about the computational complexity of a problem. Theorem 12 says MEC has a universal vulnerability to extensions by low-credence theories; Theorem 15 says the weighted Highest Median lacks that vulnerability. Those are properties of aggregation rules, not complexity classifications to transfer to a continuous population.
Theorem 12’s quantifiers matter. Its added theory can be tailored to the action set and given arbitrarily extreme evaluations. A “minimum mass needed to add a dominating theory” question therefore needs a new restriction on which theories may be added, or on their evaluations. Without one, it largely restates the theorem’s construction; with one, the restriction supplies a new problem that the paper does not define. Theorem 15 has the same problem as an anchor: it establishes a rule property, not a computational task.
The proposed Minimum Dominant-Mass Detection question avoids adding theories by searching subsets of a fixed support. But that is a new combinatorial problem, not a continuous version of Theorem 12. Its “mass” is the paper’s existing credence \(c(t)\): the paper already represents moral perspectives by weighted types and permits rational weights. If those weights are interpreted as stakeholder shares, the model is already a distribution over types; if they remain credences, they are not population counts. Either way, introducing high multiplicity does not produce a distinct input or bridge from a discrete computational result. Restricting to subsets of types also means the search is over theory classes, not over a large electorate.
A stronger proposal might let stakeholder mass move between complete evaluation types and ask for the least costly change that makes a different action win. That is coherent and close to bribery or robustness. But it adds a perturbation model and costs absent from the paper, and its distributional input is again the framework the paper already uses. It could be worthwhile as a new project; it is not a continuous mirror of a computational result here.
So the negative case is strongest under the requested anchor standard: neither theorem anchors a computational continuization, and the proposed detector already has the paper’s weighted model as its input. It is not airtight as a claim that no related computational project could be worthwhile. The mass-transfer question shows why that broader claim is harder to defend.
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.