| paper | d4-conf-ijcai-2023-0382 |
| authors | — |
| venue | ? |
| filed under | unclassified |
| judged by | gpt-6-luna / xhigh (triple__luna__xhigh__c6r1) |
| judge confidence | high |
| authors would recognise it | unclear |
The numbered items are definitions, and the paper’s complexity discussion points to results in a companion work rather than stating a named computational result here. The proposed policy optimization is a plausible extension, but it is not a computational result of this paper to mirror.
fails bit a — no named computational result to mirror
The proposed optimization would draw on the paper’s societal-harm aggregation discussion while adding policy search; it would leave the individual harm definitions, probability weighting, and benefit framework outside.
This paper has no named computational result to anchor a compliant case on: the numbered items are Definitions 1–4, not theorems about complexity or algorithms. The closest claim is in the conclusion: it says that deciding harm is DP-complete in the full paper [Beckers et al., 2022b], and that quantitative harm has completeness in a matching functional class. That claim is unnumbered and points to another work, so I cannot present it as one of this paper’s required anchors.
The strongest candidate mirror I can offer is therefore unanchored. Call it *Continuous Collective Harm Minimization*. An instance has a finite set of complete causal-utility types \(T\), rational masses \(\mu_t\) summing to \(1\), and for each type its causal model, context distribution, utility, default utility, and probability-weighting function. It also has a set of intervention variables, identifiable groups \(\mathcal G\subseteq 2^T\), and rational fairness parameters \(\alpha,\beta\). For a policy \(\mathbf{x}\), let \(h_t(\mathbf{x})\) be the weighted quantitative harm to type \(t\). Define \(\bar h(\mathbf{x})=\sum_t\mu_t h_t(\mathbf{x})\) and, for each group of positive mass, \(\bar h_G(\mathbf{x})=\sum_{t\in G}\mu_t h_t(\mathbf{x})/\sum_{t\in G}\mu_t\). The question is to find a policy minimizing
\[
\bar h(\mathbf{x})+\alpha\cdot\mathbf{1}\!\left[\exists G\in\mathcal G:\bar h_G(\mathbf{x})-\bar h(\mathbf{x})>\beta\right].
\]
A solution is a minimizing policy and its objective value.
A plausible regime is a large deployment of a standardized treatment or driving policy: many people share a small number of complete causal-risk profiles, utilities, defaults, and group memberships. Mass represents population share; the intervention remains the policy choice. This tracks the paper’s explicit concern with societal harm and disproportionate effects across groups, rather than treating its probability distributions as the continuum.
My tentative expectation is that the individual harm-computation difficulty persists even at \(\tau=1\), so population continuization would not remove the causal combinatorics. Questions include whether harm evaluation becomes tractable under useful restrictions on causal models, and how the number of types, intervention variables, contexts, and protected groups affect complexity.
The weakness is decisive for an anchor-based case: the paper does not state this optimization problem or analyze its algorithms. It studies definitions and aggregation principles, while its complexity discussion points to single-agent results in a companion paper. Thus this is a recognizable possible computational extension of the paper’s subject, but I cannot honestly claim that a numbered result in this paper establishes it as the paper’s continuous mirror.
The case against is strongest at the level of what this paper actually contributes. Its numbered results are Definitions 1–4: they define individual harm and propose ways to aggregate it. The paper does mention complexity, but it points to the full paper for the single-agent result that harm computation is DP-complete, then argues that complexity is unlikely to impede practical use. It gives no algorithmic or complexity result for computing societal harm or choosing a policy. So there is no named computational result here whose continuous version could be evaluated.
The proposed *Continuous Collective Harm Minimization* is a plausible research problem, and it tracks the paper’s stated interest in comparing interventions. But it adds a policy-search problem the paper does not formulate or analyze. The closest mirror would be to compute harm for a fixed policy over a society with masses \(\mu_t\). Once harm \(h_t\) has been computed for each complete type, the aggregate is a weighted sum, with group averages checked separately. The hard causal computation remains at the type level; population mass supplies weights. Optimizing over policies could be useful, but its difficulty would depend on a policy space and constraints supplied by a new problem formulation, not on a computational result in this paper.
There is also a mismatch around the paper’s fairness motivation. Its groups may be as small as a single identifiable person, and it emphasizes concentrated harm to one victim. In an atomless continuum, that person has zero mass, so the stated group-average test cannot represent them. Restricting attention to positive-mass demographic groups gives a sensible continuous model, but drops that individual-concentration case. A worst-case or essential-supremum criterion could retain some concentration sensitivity, but it would be a new aggregation choice; a finite set of named individuals would instead remain an exceptional atomic part of the model.
These points make the proponent’s candidate an extension of the paper’s subject, not a mirror of a result it establishes. They do not prove that no useful continuous-population study of societal harm is possible: repeated causal-risk profiles and positive-mass groups are plausible, and the proposed problem is well-posed. The negative case is therefore strong against calling this paper a computational mirror for ChoCo, but weak as a universal claim that no worthwhile mirror could be devised.
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.