| paper | d4-conf-atal-2017-0221 |
| authors | — |
| venue | ? |
| filed under | unclassified |
| judged by | gpt-6-luna / xhigh (triple__luna__xhigh__c6r1) |
| judge confidence | medium |
| authors would recognise it | unclear |
Theorem 2 is a clear computational result, and the proposed cohort problem is precisely statable. But the mirror replaces individually strategic agents with centrally coordinated cohorts and replaces an agent’s temporal goal with a population-threshold goal; neither change follows from multiplicity alone. Whether the authors would accept that organizational game as their problem remains undecided.
fails bit b-changes — the continuum limit changes the question
Cohort members need not choose the same strategy just because they share a type, and the mass-threshold objective replaces the named agent’s goal in Theorem 2.
fatal: True
It is undecided whether the authors would treat centrally coordinated cohorts with a mass-threshold goal as an influence-game winning-strategy problem. A direct author judgment on that modeling choice would settle it.
The proposed mirror covers Theorem 2’s winning-strategy existence problem under unanimous aggregation; it leaves the Nash-equilibrium complexity results and the network-structure propositions aside.
The strongest mirror I see is a type-cohort winning-strategy problem for a large, block-structured social platform. The many agents are users in a small number \(\tau\) of large cohorts: each cohort has the same initial opinions, visibility, goal, and network role. Its mass \(\mu_t\) is its fraction of the population, with rational masses summing to one. Cohorts are connected by the paper’s directed influence relation, interpreted as complete connections between the relevant blocks. Opinions remain binary, disclosure actions remain reveal or hide, updates remain unanimous, and time remains discrete. The continuous object is the population distribution; it is not the opinion or outcome space.
Here is a precise problem. Given the issues, cohort types \(T\), masses \(\mu_t\), block influence graph, initial opinions and visibility, a target cohort \(t^\star\), and an LTL goal \(\gamma^\star\), ask whether \(t^\star\) has a perfect-recall disclosure strategy that ensures \(\gamma^\star\) against every strategy of the other cohorts. A cohort strategy applies one action to all members of that cohort at each history. To make population mass operational in the goal, let \(M_{p,x}(S)=\sum_{t:B_t(p)=x}\mu_t\). For example, a goal of eventually reaching stable consensus on at least a \(q\) fraction of society, for each issue in \(J\), is
\[
\mathbf{F}\mathbf{G}\bigwedge_{p\in J}
\bigl(M_{p,0}\ge q\ \lor\ M_{p,1}\ge q\bigr).
\]
The instance includes rational \(q\); the answer is yes exactly when such a target-cohort strategy exists.
My anchor is Theorem 2, proved in this paper. It states that, under the unanimous issue-by-issue aggregator \(F_i^U\), E-WINNING\(_i(F_i^U)\) for perfect-recall strategies is in EXPTIME: given an influence game and agent \(i\), decide whether \(i\) has a strategy that achieves their goal against every strategy profile of the other agents. The paper proves the upper bound by reducing the problem to ATL model checking on the game’s concurrent game structure, using the model-checking bound cited from [1]. It does not establish EXPTIME-completeness. The cohort version keeps the strategic question—can this player guarantee their temporal goal against the others?—while expressing the population-level target in mass terms. In the block model, the same ATL construction appears to give an EXPTIME upper bound in \(\tau\), the issue count, goal size, and mass bit length: the mass predicates can be evaluated by exact rational arithmetic on each state.
For the broad-support problem, I would provisionally expect the hardness-transfer side of the trichotomy. Any finite instance embeds by giving each original agent its own type of mass \(1/n\), preserving its network position, initial state, and goal. Thus the continuous formulation includes the discrete winning-strategy problem as a special case. That is a structural reason to expect any discrete lower bound to carry over; it is not a lower bound proved by this paper. For the small-\(\tau\) regime that motivates the mirror, whether there is a genuine complexity reduction remains open. The paper’s result supplies an upper bound, not evidence that this high-multiplicity slice is easy.
This is a plausible question for these authors because their model already treats agents’ disclosure decisions, network-mediated unanimous updates, perfect recall, and temporal goals. Their references to opinion-diffusion models provide possible machinery, but the paper’s stated computational problem is about strategic solution concepts. Continuizing the population while keeping that computational question is therefore a meaningful extension. The mirror covers Theorem 2’s winning-strategy existence result under the unanimous aggregator; it does not claim to mirror the paper’s Nash-equilibrium results or its network propositions.
The weak point is the cohort player. The paper lets individually indistinguishable agents choose different strategies, whereas this mirror gives a whole cohort one common policy and uses a block network. That restriction does not follow merely from agents sharing a type. It needs a scenario in which a platform, organization, or other coordinating body actually sets disclosure policy for a cohort. If the authors do not accept cohort-level control as a natural version of their strategic-disclosure problem, this mirror loses much of its force.
Theorem 2 is a genuine computational anchor: it asks whether an agent has a perfect-recall strategy that guarantees their temporal goal, and gives an EXPTIME upper bound. The strongest objection is about whether its strategic object survives the move to population masses.
The proposed cohort mirror gives every member of a type the same action. Sharing opinions, goals, and network roles does not impose that agreement: the paper quantifies over strategies chosen by each agent, including different strategies for otherwise identical agents. The mirror also replaces the goal of a named agent with a threshold on the mass holding an opinion. Both changes may define a sensible organization-level control game, but they change the players and the objective in Theorem 2.
Allowing cohort members to choose independently preserves those choices, but creates a problem specific to this paper’s unanimous update rule. One dissenter can prevent an update. In a society of \(n\) agents, a single dissenter has mass \(1/n\); as \(n\) grows, that mass tends to zero while the dissenter can still change the transition. A distribution over types cannot treat that sequence like its limit without choosing a convention: pointwise unanimity lets even a measure-zero dissenter matter, while unanimity almost everywhere ignores such dissent and can disagree with every finite instance in the sequence. With arbitrary perfect-recall strategies, agents can also split into different behavior groups over time. The finite-type masses alone do not preserve the identity-sensitive deviations that the winning-strategy quantifier permits.
Keeping a distinguished target agent does not fully resolve the issue. The target can remain a tagged individual while the rest of society becomes continuous, but then the target has zero mass and the model must retain their strategic effect separately. That is a possible hybrid model; its dynamics still face the unanimity discontinuity among the other agents. Replacing unanimity with a mass threshold could make a more stable continuous game, and the paper itself identifies alternative aggregation procedures as future work. That would be a plausible new question, though its value would depend on whether the authors regard the changed update rule as a continuation of this result.
So the supplied cohort mirror does not establish a faithful high-multiplicity version of Theorem 2: its synchronized policies suppress deviations, and its mass goal changes what the target seeks. The more faithful versions run into the mismatch between exact unanimity and mass-based states. This is the strongest case against the anchor. It is not a decisive case against every possible mirror. A hybrid target-and-population model or a mass-based aggregation rule could still be worthwhile if the authors recognize it as their problem; I cannot honestly rule those out.
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.