Fairness in Contextual Resource Allocation Systems:

Strictly under the programme’s anchor rule, this paper has no qualifying named computational result. Propositions 1–7 are proved in the paper, but they establish incompatibility and existence conditions, not NP-hardness, polynomial-time solvability, FPT, or another complexity classification. I would therefore not claim that the paper itself proves a Class A result.

The strongest positive case is nevertheless a credible mirror of its central policy-design question, anchored—if the programme accepts an exact feasibility characterization as a broadened notion of computational anchor—on the paper’s Proposition 4, proved here.

Call the problem Minimum-Disparity CSP Allocation\(_\infty\). An instance has finite protected-group types \((g,\ell)\), where \(g\in G\) is group status and \(\ell\in L\) combines the paper’s baseline-risk and treatment-effect strata. The input gives rational masses \(\mu_{g,\ell}\), with total mass one; baseline success rates \(b_g=P(Y^0=1\mid G=g)\); treatment effects \(\tau_\ell=E[Y^1-Y^0\mid L=\ell]\); and a treatment budget \(B\).

The decision variable is a randomized allocation policy \(q_{g,\ell}\in[0,1]\), the fraction of type \((g,\ell)\) receiving treatment 1. To impose conditional statistical parity in allocation, require

\[ q_{g,\ell}=q_{g',\ell} \]

for every pair of groups having positive mass in stratum \(\ell\). Thus the policy is effectively a function of \(L\), exactly as in Proposition 4. The budget constraint is

\[ \sum_{g,\ell}\mu_{g,\ell}q_{g,\ell}\le B. \]

For each group, the expected outcome is

\[ O_g(q)=b_g+\sum_{\ell}P(L=\ell\mid G=g)\tau_\ell q_{g,\ell}. \]

The objective is to minimize the maximum pairwise outcome disparity,

\[ \min \; z \]

subject to \(|O_g(q)-O_{g'}(q)|\le z\) for all groups. The exact version asks whether the optimum is \(z=0\), i.e. whether conditional allocation parity and outcome parity can hold simultaneously.

This is a genuine population continuization. Mass represents the fraction of the service population in each complete policy-relevant type; treatment remains a finite, indivisible resource choice. The continuous object is the society and the fractional allocation across it, not merely a probabilistic outcome space. With explicitly listed finite types, the formulation is a rational linear program and is therefore an expected Class A problem. Proposition 4 supplies the exact algebraic condition characterizing when its zero-disparity version is feasible:

\[ \sum_{\ell}P(\mu_L(\ell)=1)\tau_\ell \left[P(L=\ell\mid G=g)-P(L=\ell\mid G=g')\right] = P(Y^0=1\mid G=g')-P(Y^0=1\mid G=g). \]

The paper does not state the LP or its complexity; that would be a new computational result generated by continuization, not a result already proved by the authors.

The regime is plausible in coordinated-entry housing systems. A large city may have tens of thousands of applicants, while the policy-relevant description uses a relatively small number of protected groups, standardized vulnerability scores, treatment-effect strata, eligibility categories, and resource types. A type could mean “Black applicant, score 12, eligible for PSH, treatment-effect class 3, baseline-risk class 4,” with all other information used by the policy included as well. The high-multiplicity interpretation is strongest for repeated service cohorts or administrative categories, not for a seven-person bespoke allocation committee.

The mirror should be recognizable to the authors because it preserves their objects: protected groups, legitimate features, randomized treatment policies, scarce-resource budgets, baseline outcomes, treatment effects, conditional statistical parity, and outcome parity. It does not replace their fairness question with an unrelated welfare objective. It turns their distributional statements into a finite-type policy-optimization problem.

I would not add separate anchors for Propositions 1–3 or 5–7. They are useful supporting results, especially Proposition 3’s claim that richer treatment-effect information can reduce outcome disparity and Proposition 7’s group-aware parity guarantee, but they do not define distinct computational problems. The proposed mirror covers only the joint allocation/outcome-fairness question represented by Proposition 4, not the whole framework, the empirical housing claims, or every fairness definition.

The main weakness is serious: the paper already reasons over an abstract distribution \(P\), so an opponent can say that the proposed LP merely discretizes or operationalizes an existing statistical model rather than exposing a new continuous computational object. Moreover, if each person has an idiosyncratic covariate vector or treatment effect, the number of complete types may grow with the population and the high-multiplicity gain disappears. The strongest defence is that the paper itself uses finite risk-score strata and policy summaries, and that a standardized service system genuinely contains repeated cohorts. Still, this is a qualified positive case: a convincing mirror of the paper’s central fairness-design question, but not a paper-backed computational-complexity anchor under the programme’s strictest standard.