| paper | Responsibility Gap in Collective Decision Making |
| authors | Pavel Naumov, Jia Tao |
| venue | IJCAI 2025 |
| filed under | voting · theory |
| judged by | gpt-5.6-luna / xhigh (triple__luna__xhigh__c2r1) |
| judge confidence | high |
| authors would recognise it | no |
The paper's numbered results are structural characterizations, implications, and counterexamples; none asserts an algorithmic or complexity property of a computational problem, so bit (a) fails independently of modelling. The proposed mass-based version either makes population mass decorative or replaces unilateral individual responsibility with collective responsibility. That is a re-modeling rather than a continuous population version of the paper's result.
fails bit a — no named computational result to mirror
Theorems 1–3 and the lemmas characterize gap-freeness and epistemic dictatorship, but no named computational result is covered or available to mirror.
The strongest honest positive case is only a near-miss: the paper has no qualifying computational anchor.
Its named results are structural characterisations, not complexity results. Theorem 1, proved by the authors in the full version, says that a perfect-information mechanism is gap-free iff it is an elected dictatorship. Theorem 2 proves that elected epistemic dictatorship implies epistemic-gap-freeness, and Theorem 3 proves that epistemic-gap-freeness implies elected semi-epistemic dictatorship. Lemmas 2, 3, 5, and 6 provide counterexample mechanisms. None asserts membership in \( \mathrm{P} \), NP-hardness, W[1]-hardness, FPT, or any other algorithmic-complexity classification. Thus there is no valid ChoCo anchor and, strictly speaking, no continuous problem to state “per anchor.”
The most plausible future mirror would take a finite set of agent-role types \(T\), a mass distribution \(\mu\) over those types, and a decision tree whose action and information structures are common within each type. A continuum analogue could ask whether every root-to-leaf outcome is made responsible by a positive-mass type possessing a uniform strategy, or whether the mechanism can be redesigned at minimum modification cost so that this holds. A natural regime would be, for example, a large population of interchangeable autonomous subsystems occupying a small number of safety-critical roles.
But this exposes the central weakness immediately. The paper’s responsibility predicate is individual: one agent must have a unilateral strategy. In an atomless population, changing one agent changes the aggregate action distribution by measure zero, so individual responsibility generally disappears. If responsibility is instead assigned to an entire positive-mass type-cohort, the proposed mirror becomes a theory of collective or bloc responsibility, not a continuization of the paper’s definition. If mass merely records how many duplicates of each agent-role exist while responsibility remains individual, \(\mu\) is largely decorative and the high-multiplicity relaxation has not done computational work.
So Theorem 1 suggests an interesting population-level design problem, but the paper does not supply the required named computational result, and its central notion is not naturally stable under continuization. Under the programme’s rules, this paper should therefore be recorded as having no positive continuous-mirror case, while noting the above structural near-miss as a possible direction for a different, explicitly group-responsibility paper.
This paper should be rejected as a ChoCo mirror. The proponent’s strongest point is also decisive: none of the paper’s named results is a computational-complexity result. Theorem 1 is a structural equivalence, Theorem 2 is an implication, and Theorem 3 is an inclusion. The remaining lemmas provide finite counterexamples. None specifies an input-size model or establishes polynomial-time solvability, hardness, approximation, parameterized complexity, or an optimization problem. The backward-induction definitions suggest a finite verification procedure, but that is not a stated computational result to continuize.
The proposed Theorem 1 mirror also fails at the level of the model. In the paper, an agent is an individually identifiable strategic coordinate of the choice function
\[
\tau_v:\prod_{a\in A}\Delta_v^a\to \operatorname{Ch}_v.
\]
Duplicating an agent does not have a canonical meaning: one must specify how \(\tau_v\), the information partitions, and the agents’ strategic interactions change. Identical action sets and information do not imply that two agents are interchangeable in the mechanism.
Suppose one imposes the most favourable symmetry and lets the mechanism depend only on the aggregate distribution of actions. Then individual responsibility disappears in the continuum. Changing one agent changes the aggregate by a null set, so any ordinary measurable aggregate rule gives the same outcome before and after that unilateral deviation. The paper’s sets \( \operatorname{win}_a(o) \), \( \operatorname{uwin}_a(o) \), and \( \operatorname{ewin}_a(o) \) therefore lose their intended meaning.
The obvious repair is to make a positive-mass cohort jointly responsible. But that is not a continuization of the paper’s predicate, which requires one agent to possess a strategy. It is a new theory of bloc, coalition, or organizational responsibility. Likewise, defining
\[
R_o(\mu)=\sum_{t:\,t\text{ is responsible for }o}\mu_t
\]
would produce a potentially interesting quantitative notion, but the paper’s original gap-free condition would merely become \(R_o(\mu)>0\). For a fixed support of types, the actual masses are irrelevant; only whether a type has zero or positive mass matters. Optimizing \(R_o\) would be a newly invented objective, not a mirror of Theorem 1.
Theorem 2 and Theorem 3 fare no better. Their epistemic structure concerns agent-specific indistinguishability relations over decision nodes, not population proportions. A type can include an information partition, but positive-mass duplication again either leaves individual epistemic responsibility intact as a purely support-level property or forces responsibility onto the cohort. In the latter case, the result becomes a new theorem about collective know-how, not a continuous version of elected epistemic or semi-epistemic dictatorship.
One could retain a finite set of individually responsible principals and surround them with a continuum of ordinary agents. That preserves the paper’s semantics, but the continuum then contributes no responsibility-relevant substance: all responsibility is carried by the finite atomic principals. Alternatively, one could sample a finite committee from a distribution over agent types, but that is a probability model over finite mechanisms, not a continuous society.
Thus the paper offers no qualifying computational anchor, and its central primitive is not stable under population continuization. A mass-based redesign might inspire a different paper on collective responsibility, but there is no worthwhile continuous mirror of these results within ChoCo’s population-continuization programme.
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.