| paper | d4-journals-jair-2023-0036 |
| authors | — |
| venue | ? |
| filed under | unclassified |
| judged by | gpt-6-luna / xhigh (triple__luna__xhigh__c6r1) |
| judge confidence | high |
| authors would recognise it | yes |
The allocation claim has a clear continuous analogue: a transportation LP over household types and service capacities. But the paper states it as an unnumbered Claim, while the required computational anchor is a numbered theorem, lemma, corollary, or proposition; with no qualifying anchor, bit (a) fails.
fails bit a — no named computational result to mirror
The proposed mirror covers the service-allocation claim in §5.1.1; it leaves the BART estimation, fairness analyses, and dynamic matching questions outside the mirror.
The strongest positive case is the paper’s unnumbered “Claim” in §5.1.1. It proves that assigning each household exactly one service, subject to service capacities and minimizing expected re-entries, is equivalent to maximum-weight bipartite \(b\)-matching: give each household–service edge weight \(1-p_{ij}\), household capacity \(1\), and service capacity \(C_j\). The paper proves this equivalence and concludes that the allocation is solvable in polynomial time.
There is a formal limitation: the supplied paper has no numbered Theorem, Lemma, Corollary, or Proposition asserting a computational result. So this Claim is the closest real anchor, but it does not meet your requested numbering convention; I can’t supply a number that is not printed.
A continuous mirror of that Claim is Continuous Counterfactual Service Allocation. An instance consists of a finite set of household types \(T\), each with population mass \(\mu_t\), a set of services \(J\), a predicted re-entry probability \(p_{tj}\) for assigning type \(t\) to service \(j\), and capacity \(\kappa_j\) for each service. Types are complete for this optimization: households of type \(t\) have the same eligibility and predicted probability for every service. Choose masses \(x_{tj}\geq 0\) to assign, minimizing expected re-entry mass
\[
\sum_{t\in T}\sum_{j\in J} p_{tj}x_{tj},
\]
subject to
\[
\sum_{j\in J}x_{tj}=\mu_t \quad (t\in T), \qquad
\sum_{t\in T}x_{tj}\leq \kappa_j \quad (j\in J),
\]
and \(x_{tj}=0\) for ineligible type–service pairs. A solution is the assignment mass \(x\). Here mass means a fraction of the cohort; the fractional assignment is meaningful because each type can contain many households.
The plausible regime is a large city or regional homeless system allocating services to large weekly or monthly intake cohorts. Many households can share the same recorded profile, eligibility, and counterfactual risk vector, while the number of such distinct types remains modest relative to the cohort size. The paper itself studies weekly batches and service capacities, so this is close to its allocation question: the continuous model retains the choices, predictions, and scarcity constraint that drive the result. Its expected classification is tractable: it is a transportation LP, and the paper’s discrete \(b\)-matching result is strong evidence that the authors would recognize the underlying allocation problem.
This covers the service-allocation result, not the paper’s BART estimation, fairness analysis, or dynamic matching questions. No second anchor is needed: the paper has no comparable named hardness result, and stretching its fairness discussion into another anchor would make the case less precise.
The weakest point is that the paper’s proved result is already polynomial-time and concerns individually identified households; the continuous formulation mostly aggregates that same allocation problem. Its value for this programme would be to make the high-multiplicity version explicit and characterize it, not to claim a dramatic complexity change. The additional questions are whether the type count stays small in realistic cohorts, how prediction uncertainty should enter \(p_{tj}\), and how to impose the paper’s per-household harm bound using types that also record each household’s original service.
The only strong computational anchor is the Claim in §5.1.1: the service assignment can be solved as weighted bipartite \(b\)-matching. It is a real computational result, even though it is unnumbered. Its natural mirror is also clear: group interchangeable households into types and assign mass \(x_{tj}\) to services, minimizing \(\sum_{t,j}p_{tj}x_{tj}\) subject to type-mass and service-capacity constraints. The paper’s per-household harm constraint could be preserved by including original service in the type.
The best objection is about whether this is a meaningful high-multiplicity regime for the problem the paper actually solves. It optimizes separate weekly cohorts: 10,043 households over 166 weeks, or about 60 households per cohort on average. Its capacities are estimated from the number who actually entered each service that week. The large total population therefore does not supply a large batch at the decision point. To define types faithfully, one would need to retain the covariates that drive each household’s counterfactual probability, eligibility, and—if modeling the fairness constraint—original service. With 35 features, including continuous features, the resulting support may be close to the number of households. That is a distribution written over people, not a useful compression into repeated types.
Pooling weeks could create larger masses, but it changes the allocation problem: it suppresses the arrival timing and week-specific capacities that define the paper’s policy. A time-expanded mass model could preserve them, but then it reproduces the same small, independent weekly assignments. The paper does mention dynamic matching as a possible extension, but that would be a new question rather than a continuous mirror of this Claim.
This is the strongest case against the anchor, and it remains limited. Grouping by predicted risk vectors is a sensible alternative, and larger systems or coarser profiles could have substantial repeated mass. The paper’s result is also about allocation, not just prediction, so an aggregate model could matter. I would argue that the Claim offers little evidence of a worthwhile continuization *for this paper’s actual decision regime*, but I cannot rule out a convincing larger-cohort scenario. The negative case is therefore weak against the universal burden.
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.