| paper | d4-conf-aaai-2024-2227 |
| authors | — |
| venue | ? |
| filed under | unclassified |
| judged by | gpt-6-luna / xhigh (triple__luna__xhigh__c6r1) |
| judge confidence | high |
| authors would recognise it | no |
The paper has no numbered computational result to anchor a continuous computational question. The proposed type-level bargaining model also changes who proposes and accepts allocations, so it does not supply a recognized mirror of the paper’s individual negotiation process.
fails bit a — no named computational result to mirror
The proposed model targets the explanation-existence Proposition; it leaves the backward-induction claims, human studies, and noise experiments without a valid computational mirror.
The strongest positive case is a high-multiplicity mirror of the paper’s counterfactual-explanation guarantee. There is an anchor problem: the paper’s “Proposition”—printed without a number—says that for an AITA explicable allocation \(o\) and a human’s counterfactual \(o_i\), an explanation exists: a negotiation tree under true costs reaches an allocation \(\hat o_i\) with \(C_i(\hat o_i)\ge C_i(o)\). The proposition is proved here, by contradiction.
That is the closest anchor, but it does not meet your numbered-anchor requirement, and it asserts no P, NP-hardness, or other complexity classification. The paper also gives an unnumbered Lemma and Corollary about backward induction and acceptable allocations, but neither supplies such a classification. So this can only be a provisional positive case, not a fully compliant anchored one.
A plausible regime is a large organization assigning repeated shifts or task bundles among many workers who share a small number of complete profiles: the same task costs, capacity, and negotiation role. There, the number of workers \(N\) could be much larger than the number of types \(\tau\). A type’s mass \(\mu_k\) is its fraction of the workforce; task mass is the fraction of repeated workload assigned to it. This preserves the paper’s allocation-and-contestation question, while making the population itself continuous.
One precise candidate is Continuous High-Multiplicity Counterfactual Refutation. An instance has worker types \(k\in K\), with rational masses \(\mu_k>0\) summing to one; task classes \(j\in J\), with rational workload masses \(b_j\); per-capita capacities \(u_k\); and rational costs \(c_{kj}\). A feasible aggregate allocation is a matrix \(x\) in
\[ X=\left\{x\ge 0:\ \sum_k x_{kj}=b_j\ \forall j,\quad \sum_j x_{kj}\le u_k\mu_k\ \forall k\right\}. \]
Here \(x_{kj}\) is task mass of class \(j\) assigned to worker type \(k\), and type \(k\)’s per-capita cost is
\[ C_k(x)=\frac{1}{\mu_k}\sum_j c_{kj}x_{kj}. \]
Let AITA’s team-performance cost be a rational linear function \(P(x)\). The instance also gives a baseline proposal \(\bar x\), a counterfactual \(y\), and a type \(i\) contesting it, with \(C_i(y)<C_i(\bar x)\). Use the vertices of \(X\) as the finite offer set; on each turn, the proposer offers their lowest-cost unoffered vertex, breaking ties lexicographically. Worker types accept an offer exactly when their cost is at most their cost at \(\bar x\); AITA uses \(P\) for its offer ranking. If all offers are rejected, the last is imposed. The question is to compute the subgame-perfect continuation from \(y\), returning its offer-and-response path and terminal allocation \(\hat x\). A solution is a finite explanation path ending with \(C_i(\hat x)\ge C_i(\bar x)\).
This mirrors the Proposition’s contrastive guarantee, not the paper’s human-study claims or its noise experiments. For this version I would tentatively expect continuum-specific hardness: backward induction may have to reason over exponentially many vertices even when the type and task descriptions are compact. That is a conjecture about the proposed problem, not a result of the paper. Useful follow-up questions are whether the guarantee still holds for aggregate allocations, whether a path can be computed by optimization or separation rather than vertex enumeration, and how explanation length scales with \(|K|\) and \(|J|\).
The soft spot is substantial: the paper’s bargainers are a small set of individual humans, whereas this mirror treats whole worker types as negotiating players. The vertex-offer rule also excludes interior allocations from the bargaining protocol. Those are defensible choices for a first formalization, but the authors could reasonably reject either as changing their game. And because the paper proves no complexity classification, this anchor does not establish that continuization dissolves—or creates—the computational difficulty.
I would vote against this as a mirror of the paper’s results. The proponent’s anchor is an unnumbered existence proposition: for the paper’s finite bargaining model, an explanation exists. The paper gives no complexity classification for computing one. Its backward-induction procedure searches an allocation-enumeration tree that can contain \(n^m\) outcomes, but it makes no tractability claim. Asking for a continuous existence guarantee would extend a property of the model; asking for an algorithm to compute a continuation would be a new computational question, not a continuization of a computational result in this paper.
The strongest repair is to ask for the subgame-perfect continuation in the proponent’s aggregate model, rather than merely whether an explanation exists. But that model gives each worker type a single bargaining voice. In the paper, a particular human contests an allocation, evaluates their own cost, and may reject an offer; each human also has a turn in the negotiation. An aggregate \(x\) records how much work goes to a type, not how it is distributed among its members. Two allocations with the same \(x\) can leave the contesting person with different workloads and costs. The mass model therefore cannot, by itself, produce the paper’s personalized explanation.
Keeping every worker as a separate bargainer preserves that feature, but then the type distribution does not encode who proposes, who objects, or whose acceptance is required. Merging identical workers into a bloc restores a finite game only by changing the players and the veto: a type’s representative now speaks for the group. That may be a sensible model of union or cohort bargaining, but the paper studies individual teammates contesting an AI allocation. Restricting offers to vertices of \(X\) makes the proposed game finite; it does not bridge that change in agency.
A repeated-shift workforce with worker cohorts and elected representatives is the strongest scenario for the mirror, and it makes the negative case less than airtight. It could be a worthwhile neighboring problem. The reason to withhold the green verdict is that it needs a different bargaining institution to make population mass computationally usable; the paper’s own central result does not supply a computational claim that survives that change.
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.