| paper | Pushing the Limits of Fairness in Algorithmic Decision-Making |
| authors | Nisarg Shah |
| venue | IJCAI 2023 |
| filed under | frontier · tools-data |
| judged by | gpt-5.6-luna / xhigh (triple__luna__xhigh__c2r1) |
| judge confidence | high |
| authors would recognise it | yes |
The paper’s only numbered statements are Definitions 1–5; it contains no numbered theorem, lemma, corollary, or proposition asserting a computational result. The block-template peer-review mirror is a coherent continuous population question, but its plausibility cannot repair the missing computational anchor required by bit (a).
fails bit a — no named computational result to mirror
The proposed mirror covers only the unnumbered §3.3 peer-review claim and leaves the paper’s fairness definitions, survey applications, and externally cited results untouched.
The strongest honest answer has an important qualification: the supplied paper contains no qualifying named computational result. Its only numbered statements are Definitions 1–5. There is no numbered Theorem, Lemma, Corollary, or Proposition asserting tractability, hardness, or any other complexity classification. The closest passage is the unnumbered claim in §3.3 that a core reviewing assignment “always exists, and can be computed efficiently,” attributed to Aziz, Micha, and Shah (2023). It is therefore neither a numbered anchor nor a result proved in this paper. Strictly under the programme’s rules, there is no valid anchor and hence no fully compliant positive case.
If the reader permits that explicit unnumbered claim as a provisional anchor, the strongest mirror is the peer-review application, not the paper’s already-divisible-goods model. Call it Continuous Core Peer Review, \(\mathrm{CCPR}_\infty\).
An instance has a finite set \(T\) of complete researcher types and rational masses \(\mu_t\), with \(\sum_{t\in T}\mu_t=1\). A type specifies a paper profile, reviewer profile, reviewer capacity, compatibility constraints, and a rational value \(v_t(r)\) for receiving a review from type \(r\). Think of a large conference in which many papers and reviewers recur by stable topic, expertise, institutional, and conflict classes, so \(n\) is very large while \(\tau=|T|\) is moderate. Each unit of type \(t\) contributes one unit of review capacity and represents one paper requiring one review; this is the paper’s single-author setting in mass form.
The decision variable is \(x_{r,t}\), the mass of type-\(r\) reviewers assigned to type-\(t\) papers. It must satisfy
\[ x_{r,t}\ge 0,\qquad x_{r,t}=0\text{ on incompatible pairs}, \]
\[ \sum_t x_{r,t}=\mu_r,\qquad \sum_r x_{r,t}=\mu_t. \]
The utility of type \(t\) is
\[ u_t(x)=\frac{1}{\mu_t}\sum_r v_t(r)x_{r,t}. \]
A coalition is a nonzero mass vector \(\lambda\) with \(0\le\lambda_t\le\mu_t\). It blocks \(x\) if it can internally reassign its own reviewing capacity, through some compatible \(y\), with
\[ \sum_t y_{r,t}=\lambda_r,\qquad \sum_r y_{r,t}=\lambda_t, \]
such that
\[ \frac{1}{\lambda_t}\sum_r v_t(r)y_{r,t}\ge u_t(x) \]
for every participating type, with strict inequality for at least one type. A solution to \(\mathrm{CCPR}_\infty\) is a feasible assignment \(x\) with no blocking coalition; under the same mild preference and load assumptions used in the paper’s peer-review claim, one asks for such an \(x\), or a certificate that none exists.
This is a credible population continuization. The mass represents many interchangeable researchers, not fractional treatment of a named researcher. Rational masses can be expanded into cloned finite instances, while a symmetric assignment of those clones induces the matrix \(x\). The entitlement is exactly the paper’s entitlement: a group may use the reviewing capacity contributed by its members to improve the treatment of its own papers. The mirror changes the population representation, while preserving the paper’s matching, preferences, capacity, conflict, and core-deviation structure.
I would expect the promised version to be Class A. Once \(x\) is fixed, detecting a blocking coalition is naturally a finite linear program in \(\lambda\) and \(y\); constructing \(x\) is plausibly a transportation, assignment-game, or weighted Nash-welfare computation over the type system. The unrestricted model may instead become continuum-specifically hard, especially with arbitrary compatibility and idiosyncratic preferences. Further questions include exact versus \(\varepsilon\)-core computation, rounding guarantees for finite conferences, multiple reviews per paper, and whether individual self-conflicts disappear legitimately as measure-zero diagonal constraints.
The weakest point is decisive: this mirror is anchored only to an unnumbered survey claim cited from elsewhere, so it cannot meet the paper-selection rule as written. There is also a substantive modelling risk if real peer-review conflicts are individual-specific; encoding every such relation into the type makes \(\tau\) approach \(n\), destroying the high-multiplicity regime. Thus the paper supports a plausible continuous question, but supplies no valid numbered computational anchor for it.
The decisive negative is that this paper supplies no qualifying computational anchor. Its only numbered statements are Definitions 1–5; there is no numbered theorem, lemma, corollary, or proposition about tractability, hardness, approximation, or algorithms. The peer-review sentence that an assignment “can be computed efficiently” is an unnumbered survey claim attributed to Aziz, Micha, and Shah (2023), not a result established in this paper. Treating it as an anchor changes the task from continuizing this paper to inventing a new problem from its bibliography.
Even granting that provisional anchor, the proposed \(\mathrm{CCPR}_\infty\) does not faithfully preserve the peer-review object without imposing a severe symmetry assumption. A researcher’s relevant type must include the submission, reviewer capacity, preferences, and compatibility or conflict neighborhood. In actual peer review, those neighborhoods are typically individual-specific. Exact aggregation therefore gives \(\tau=\Theta(n)\), destroying the moderate-type high-multiplicity regime. If one aggregates researchers merely by topic or institution, one loses the conflict and assignment structure that makes the cited peer-review result meaningful.
The best rescue is a deliberately constructed conference that is a blow-up of a fixed finite template: many interchangeable papers and reviewers in each block, with block-constant preferences and conflicts. That is mathematically coherent, but it is a new fractional two-sided-market/core problem, not a mirror of a named computational result in this paper. The proponent’s claims that coalition detection is an LP and that construction is “plausibly” a transportation or Nash-welfare computation are research suggestions, not results supplied by the paper.
This is not the mistaken objection that divisible goods already make the paper continuous: that would concern the outcome space, not the population. Nor is existing high-multiplicity work a collision. The narrower point is that the paper is an axiomatic survey whose computational content is borrowed, unnumbered, and external. I would therefore reject it as a source for ChoCo’s continuization programme.
The honest limitation is that, if the anchor rule is relaxed, the block-template peer-review model is a plausible independent ChoCo problem. The universal claim that no worthwhile scenario exists cannot be defended on substantive modelling grounds alone; the negative case succeeds because this paper contains no valid computational result to 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.