| paper | d4-conf-aaai-2016-0673 |
| authors | — |
| venue | ? |
| filed under | unclassified |
| judged by | gpt-6-luna / xhigh (triple__luna__xhigh__c6r1) |
| judge confidence | high |
| authors would recognise it | no |
Theorem 2 and Theorem 5 are computational hardness results, but the proposed masses range over advocates assigned to separate games. This computes a weighted share of instances with a successful move or strategy; the paper’s game itself has two strategic roles, and its arguments are interdependent components rather than agents whose mass can vary. Neither anchor therefore yields an in-scope continuous-population question.
fails bit b-none — no population with multiplicity to relax
The proposed caseload is a distribution over independent game instances, so its mass aggregates which instances are solvable rather than changing the population inside the strategic argumentation problem.
fatal: True
The proposed portfolio covers the grounded Desired Outcome problem for P and grounded espionage strategies; it leaves stable-semantics aims, spoiler variants, and concrete-language results aside. The portfolio construction does not provide an in-scope continuous-population mirror for the covered results.
The most defensible mirror is a high-multiplicity portfolio of independent disputes, rather than one dispute played by a continuum of people. Imagine a legal-aid service handling a very large caseload of recurring, templated matters. Each agent is a proponent-side advocate assigned one dispute. Their type records the complete game they face: the split argumentation framework, target argument, available arguments, information, and aim. Advocates of the same type are exact copies in every respect the game uses. If there are \(N\) advocates but only \(\tau\) distinct templates, type \(t\) has mass \(\mu_t=n_t/N\), given as a binary rational. This is plausible when many cases share a small set of argument structures; it is not a claim that every legal dispute has high multiplicity.
For a type \(t\), let \(I\subseteq A_P^t\) be the set of arguments its advocate plays in the move considered, and let \(\operatorname{Acc}_t(I)\) equal \(1\) exactly when the target is in the resulting framework’s grounded extension. The lead mirror is Grounded Desired-Outcome Coverage\(_\infty(q)\). Its instance is a finite-support distribution \(\mu\) over such types and a rational threshold \(q\in[0,1]\). Choose masses \(x_{t,I}\geq 0\), with \(\sum_I x_{t,I}=\mu_t\) for every type, to maximize
\[
\sum_{t,I}x_{t,I}\operatorname{Acc}_t(I).
\]
The question is whether the maximum is at least \(q\). A solution is an allocation of each type’s advocate mass among ordinary legal moves; no advocate plays a fractional argument.
The anchor is Theorem 2, proved in this paper: under grounded semantics, the Desired Outcome problem for \(P\)—given a split argumentation framework and \(P\)’s desired outcome, is there a subset \(I\subseteq A_P\) whose addition makes the distinguished argument accepted in the grounded extension?—is NP-complete. The paper proves this by reduction from SAT. The mirror’s one-type, mass-one, \(q=1\) case is exactly that problem. For the finite-support coverage version, the answer is NP-complete too: a certificate can select types whose masses meet \(q\) and give one successful move for each. I expect hardness transfer (Class B): the hard combinatorics sit inside a single argumentation framework, not in the number of advocates. The exponential-choice optimization view makes clear what would need to change for tractability: the search for a successful move, or pricing, would need a tractable structure.
A second, distinct mirror is Grounded Espionage-Strategy Coverage\(_\infty(q)\). Its types additionally record that \(P\) knows the opponent’s available arguments. For each type, allocate advocate mass \(y_{t,s}\) among strategies \(s\), where a strategy maps the public history and \(P\)’s available arguments to the next move. Let \(W_t(s)=1\) exactly when every valid sequence of opponent moves consistent with \(s\) ends in a win for \(P\). Maximize \(\sum_{t,s} y_{t,s}W_t(s)\), subject to \(\sum_s y_{t,s}=\mu_t\), and ask whether the result is at least \(q\). A solution assigns a winning policy to at least the required mass of advocates; each assigned advocate still plays an ordinary strategy in an ordinary game.
Its anchor is Theorem 5, also proved in this paper: under grounded semantics, the Winning Strategy problem—whether \(P\) has a strategy that wins against every valid sequence of \(O\)’s moves—is PSPACE-complete. The paper’s hardness construction models quantified Boolean formulas. Again, a one-type population with mass one and \(q=1\) recovers the paper’s question exactly; the finite-support threshold version remains PSPACE-complete. I expect hardness transfer for the same reason: even a single template retains the original strategy problem. Further questions include whether repeated templates admit compact shared policy representations, and whether restrictions on the attack graph yield tractable pricing or useful parameterized algorithms.
These mirrors cover the paper’s grounded Desired Outcome problem for \(P\) and its grounded espionage problem. They do not cover stable-semantics aims, the opponent’s spoiler variants, or the concrete-language extension results. Their population is the distribution over advocates assigned to repeated games; the argumentation semantics inside every case stays unchanged.
The soft spot is substantial: this batches independent games instead of putting many agents into one shared argumentation framework. The authors could reasonably call it a high-multiplicity deployment problem, rather than a continuum version of a single strategic-argumentation game. The case for accepting it is that large services really do ask what fraction of a templated caseload admits a successful move or a guaranteed strategy, and the mirror preserves the exact question for each advocate. If the programme’s gate requires one shared game whose own society is continuous, I do not see a comparably honest mirror here.
The paper has named computational results, so the strongest negative case is not that it lacks anything to mirror. The problem is what those results compute: one proponent and one opponent play one shared argumentation game. The arguments are vertices in that game, not members of a population.
Theorem 2’s proposed coverage problem is well-defined, but its optimum is just
\[ \sum_t \mu_t\, \mathbf{1}\!\left[\exists I\subseteq A_P^t: a_t\in \operatorname{GE}(A_{\mathrm{Com}}^t\cup I)\right]. \]
Each type’s advocates face separate games. If one move succeeds, all its mass can be assigned that move; if none succeeds, none of its mass contributes. The threshold question therefore measures the weighted share of instances whose Desired Outcome problem is yes. With one type and \(q=1\), it is exactly Theorem 2. With several types, it batches instances of Theorem 2. That may be useful for a legal service deciding what fraction of a templated caseload can be advanced, but the continuum is over a caseload of problems, not over participants in the game the theorem studies.
Theorem 5 has the same limitation. Its portfolio version measures
\[ \sum_t \mu_t\, \mathbf{1}\!\left[\exists s_t\ \forall\text{ valid opponent plays, }P \text{ wins under }s_t\right]. \]
Each advocate gets a separate policy for a separate game. The opponent’s moves, the private argument sets, and the histories remain local to that game. Requiring a common policy across different frameworks would instead pose a policy-synthesis or generalization problem that the paper does not define. Adding a pooled staff budget or shared resources could couple the cases, but that would make a new portfolio-allocation problem, not continuize the paper’s espionage question.
A single-game population mirror faces a more basic obstacle. The model has two strategic roles and moves consist of adding whole arguments. Its semantics depends on which arguments exist and which attack which; it has no notion of argument support mass. Duplicating identical advocates either collapses into the existing player if their choices can be pooled, or requires a new turn protocol if they cannot. Giving arguments masses would require a weighted semantics: under the paper’s semantics, any positive mass that makes an argument present gives it the same status, while a zero-mass individual contributes no argument. That is a substantial new model of argumentation, not a population limit of the game analyzed here.
So I would reject both anchors as continuous-population mirrors of this paper. The proposed caseload question is coherent and might interest legal-service operators; it is also a distributional analysis over independent instances, a construction that could be attached to almost any decision problem. The negative case is strongest on that mismatch. It cannot rule out the authors finding such an operational question worthwhile, but the paper itself gives little reason to treat it as a continuous society inside strategic argumentation.
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.