| paper | d4-conf-aaai-2023-0052 |
| authors | — |
| venue | ? |
| filed under | unclassified |
| judged by | gpt-6-luna / xhigh (triple__luna__xhigh__c6r1) |
| judge confidence | high |
| authors would recognise it | yes |
Proposition 4 characterizes when two fairness conditions hold, but the paper does not state or use an algorithmic result that the proposed feasibility problem could mirror. The rational type-mass LP is a coherent policy-design follow-up, yet it adds a computational question to the paper rather than continuizing one of its computational results.
fails bit a — no named computational result to mirror
The proposed feasibility question builds on Proposition 4's conditional allocation parity and outcome parity condition; it does not cover Propositions 1–3 or 5–7, which also state fairness characterizations or existence claims without computational guarantees.
The strongest positive opening is policy design for a large housing-services cohort with many clients but relatively few distinct profiles. Define a type by the features the policy uses—protected group, baseline-risk level, and treatment-effect level—and let \(\mu_{g,\ell}\) be the fraction of the cohort in group \(g\) and profile \(\ell\). Each client receives either treatment or no treatment; the decision variable \(q_\ell\in[0,1]\) is the treatment probability for everyone of profile \(\ell\). A concrete mirror question is: given rational type masses, baseline outcome rates, treatment effects, and a resource budget \(b\), is there a vector \(q\) such that \(\sum_\ell q_\ell\Pr(L=\ell)\le b\), treatment assignment is conditionally statistically fair across groups, and outcome rates are equal across groups? A solution is such a vector \(q\), which can be implemented by random assignment among clients of each type. This is a finite linear-feasibility problem.
The nearest named result is Proposition 4, proved in this paper (with its proof in the supplementary materials). It says that an allocation policy based on \(L\) satisfies conditional statistical parity in allocation and statistical parity in outcomes if and only if, for every pair \(g\ne g'\), \(\sum_{\ell}\Pr(\mu_{\mathcal L}(\ell)=1)\tau_\ell[\Pr(L=\ell\mid G=g)-\Pr(L=\ell\mid G=g')]=\Pr(Y^0=1\mid G=g')-\Pr(Y^0=1\mid G=g)\). That condition supplies the algebra behind the proposed feasibility question and makes the mirror recognizable: it asks whether a resource-constrained policy can meet the two fairness goals the paper studies.
But Proposition 4 is not a qualifying computational anchor. It characterizes when two fairness conditions hold; it does not classify the complexity of a computational problem or present an algorithmic result. Nor do the other numbered propositions: they give incompatibility, disparity, or existence claims, not complexity results. So the paper contains no named computational result of the kind your instructions require, and I cannot honestly present this LP as a mirror of one of its computational anchors. The proposed problem is an author-recognizable computational extension, not an anchored case.
The regime is plausible for recurring, large coordinated-entry systems: many clients enter across a region or year, while the policy groups them into a much smaller set of profiles. The weak point is substantial: the paper already describes populations through probability distributions and randomized policies, so a mass-based formulation may partly restate its existing population model. Without a named computational result, the positive case remains a credible direction for follow-up work, rather than a strong continuous mirror of a result in this paper.
The proposed anchor fails the requested test at the outset: the paper states no computational complexity, algorithm, or optimization theorem. Proposition 4 is an if-and-only-if characterization of when two fairness conditions hold. The other propositions give characterizations or existence and disparity claims. None provides a computational result to mirror.
The proponent’s policy question can be made precise. Let \(p_{g,\ell}\) be the mass of group \(g\) at legitimate profile \(\ell\), and let \(q_\ell\in[0,1]\) be a common treatment probability across groups at that profile. Add a per-capita resource budget, and ask whether the allocation-fairness and outcome-parity constraints can all hold. With finite rational input, those constraints are linear, so this is a finite LP. It is a plausible policy-design extension, and its tractability could be useful. It still asks a new computational question; it does not mirror Proposition 4’s complexity result, because that proposition has none.
The stronger mirror also runs into the paper’s own model. The paper already represents the population through a joint distribution over covariates, group status, treatments, and potential outcomes. It defines randomized policies conditionally on covariates and expresses fairness through population probabilities. In a finite-profile version, the masses \(p_{g,\ell}\) are simply those probabilities written as type masses. Continuization adds no new population representation. If the covariate space is genuinely continuous, a complexity question needs an input model for the distribution and its integrals; discretizing it returns to the finite LP. Requiring exact realized allocations or adding learning assumptions could create different computational problems, but those would add requirements absent from the paper’s fairness results.
There are plausible large-cohort settings here, and a computational follow-up might be worthwhile. The honest negative case is strongest against calling it a continuous mirror of this paper: there is no named computational result to anchor, and the paper already formulates its fairness framework distributionally. The broader claim that no related computational project could ever be worthwhile is harder to establish, but it does not rescue the proposed mirror.
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.