| paper | d4-conf-wine-2025-0015 |
| authors | — |
| venue | ? |
| filed under | unclassified |
| judged by | gpt-6-luna / xhigh (triple__luna__xhigh__c6r1) |
| judge confidence | high |
| authors would recognise it | no |
Theorem A.1 is a genuine computational result, but the proposed Mass-EFR question changes its fairness test from individual envy to envy between cohort means. That change is not forced by representing the population as a distribution over types, so this formulation does not pass the authors-recognition gate for a continuous analogue.
fails bit b-changes — the continuum limit changes the question
The model tests envy between cohort averages; the paper requires each individual to compare their bundle with every individual's bundle in a counterfactual allocation.
fatal: True
The only anchor is Theorem A.1, whose NP-completeness result is not mirrored by the cohort-average fairness test. The paper's existence, computation, tightness, and EFR-plus-PO results are not covered.
The positive case I would defend is a mass version of Theorem A.1, the paper’s hardness result. It is proved in Appendix A: given a mixed-manna instance with additive valuations, a specified allocation \(A\), and a positive integer \(k<n-2\), deciding whether \(A\) is EFR-\(k\) is NP-complete. The paper proves hardness even when all agents have identical valuations, by reducing from PARTITION. That is a useful anchor because it asks whether a *given allocation* has the paper’s common-reallocation fairness property.
A plausible high-multiplicity regime is a large public-service workforce receiving recurring benefits and duties: for example, housing placements alongside unpleasant shifts. Workers with the same complete valuation vector are one type; the supply of each standardized category scales with the workforce. The population might have \(N\) workers and \(\tau\) valuation types, with \(N\gg\tau\). This resembles the paper’s mixed-manna setting while giving population shares and per-capita supplies a natural meaning.
Here is the continuous problem I would hand to a prover. Call it Mass-EFR Verification. An instance has a finite set \(T\) of complete valuation types, rational masses \(\mu_t>0\) summing to \(1\), \(m\) item categories with rational per-capita supplies \(q_j\), additive rational values \(v_{tj}\), a baseline allocation \(x_{t,B}\) over integral bundles \(B\in\mathbb Z_{\ge0}^m\), and a reserve threshold \(\rho\). The allocation must satisfy \(\sum_B x_{t,B}=\mu_t\) and \(\sum_{t,B}x_{t,B}B_j=q_j\). Thus each unit of population gets an integral bundle, while its type’s average bundle can be fractional.
For each positive-mass cohort \(p=(t,B)\), let \(\alpha_p=x_{t,B}\) and \(z_{pj}=\alpha_pB_j\). A certificate consists of one shared reserve vector \(r\in\mathbb Q_{\ge0}^m\) and, for every cohort \(p\), a counterfactual aggregate allocation \(z^{(p)}\). Each counterfactual preserves cohort masses and item supplies. It makes cohort \(p\) envy-free in average utility: its mean bundle is at least as valuable to type \(t\) as every other cohort’s mean bundle. The mass of category \(j\) moved in that counterfactual is
\[
d_j^{(p)}=\frac12\sum_a\left|z^{(p)}_{aj}-z_{aj}\right|.
\]
Require \(d_j^{(p)}\le r_j\) for every \(p,j\), and \(\sum_j r_j\le\rho\). The question is whether such a certificate exists. The solution gives the counterfactuals and the common reserve; every mean bundle in them can be implemented as a distribution over integral bundles.
I expect this version to be tractable. Its constraints are linear: introduce variables for each counterfactual allocation and for the absolute-value bounds, then impose the envy and shared-reserve inequalities. The LP has polynomial size in \(m\), the number of positive-mass cohorts, and the input bit length. The continuous version of the finite threshold is \(\rho=k/N\), the number of reserved item copies per capita. In the Partition construction, replicated copies of each item category can be divided as mass among cohorts, so the whole-item subset choice that carries the finite hardness becomes linear allocation of mass.
The authors have some reason to recognize this as a relevant regime: their related-work discussion includes results for few types of agents. Still, the weak point is substantial. Their EFR guarantee is about each individual’s bundle versus every other individual’s bundle. Mass-EFR instead compares cohort averages. That is a change to the fairness criterion, not merely a change in notation. I think it is defensible for institutions that assess fairness across large, stable cohorts, and it retains integral bundles for each person and the paper’s shared-reserve idea. Whether the authors would accept *average cohort envy* as their continuous analogue is the gate.
This case mirrors Theorem A.1 alone. It generates two useful follow-up questions: whether finite EFR certificates converge to this mass version with a controlled rounding error, and whether NP-hardness returns when fairness is required person by person rather than at cohort-average level.
Theorem A.1 is a genuine computational anchor: the paper proves NP-completeness for deciding whether a specified allocation is EFR-\(k\). The workforce scenario is also plausible. My objection is to whether the proposed Mass-EFR model mirrors that result.
In the paper, for each individual \(i\), there must be a counterfactual allocation in which \(i\) envies no one. The LP instead compares cohort means. Those conditions can disagree even when every person still receives an integral bundle: give half of each of two same-type cohorts a good and half nothing. Both cohorts have the same mean bundle, so the LP declares their average utilities equal. Yet each person who received nothing envies someone who received the good. Realizing a mean as a lottery over integral bundles does not fix this: it preserves the distribution of bundles while the LP tests only their averages. Mass-EFR is a coherent ex ante cohort-fairness question, but it changes the paper’s fairness criterion.
A more faithful version would track bundle distributions and preserve individual envy comparisons. That runs into a scaling problem. With \(N\) people and per-capita item supplies, a fixed reserve of \(k\) items has normalized mass \(k/N\), which vanishes. Scaling the reserve to \(\rho N\) keeps positive mass, but each target-specific counterfactual can route that reserve to one individual, whose attainable bundle then grows with \(N\). Requiring the reserve to change a positive-mass cohort avoids that concentration, but returns to a cohort-level fairness condition. The proposed finite-rounding and per-person-hardness follow-ups do not resolve this mismatch; they address different versions of the fairness requirement.
That is the strongest case against this mirror: the LP may be tractable and useful, but its fairness guarantee does not preserve the result’s individual-level content. I would not green Mass-EFR as a direct continuization of Theorem A.1.
I would not, however, claim that *no* worthwhile mirror exists in any scenario. The workforce setting is credible, and the authors might recognize ex ante cohort envy as a worthwhile extension. The negative case is strongest against this proposed mirror’s fidelity; it remains weak at the programme’s recognition gate.
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.