{"id":"700d30eb-f121-453c-8728-14979019db55","arxiv_id":"1908.07784","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Power-index values over the extensions of a Dung semantics define a total pre-order on arguments, implemented in the ConArg tool.","lead":"The paper implements a ranking-based semantics for abstract argumentation that ranks arguments with cooperative game-theory power indexes (Shapley, Banzhaf, Deegan-Packel, Johnston), using Dung's classical semantics as a parameter. It reports which standard ranking properties the Shapley-based version satisfies and demonstrates the tool on a small argumentation graph.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The Deegan-Packel and Johnston formulas (Eqs. 3 and 4) are not well-formed as printed, so the rankings in Tables III and IV cannot be verified as implementations of the named indexes; this undercuts the parametric-family claim, though the Shapley-based results are unaffected.","rationale":"The reader's weakest assumption points to exactly this problem: the Deegan-Packel and Johnston formulas are malformed as printed, and the rankings reported for those indexes may not correspond to the named notions. My independent reading of Section II-B and Table IV confirms that the concern is real and load-bearing for any claim involving ρ and γ. The Shapley-based construction is standard and the numerical values in Table I are plausibly correct, so the paper's core Shapley property profile is not destroyed. The issue is fixable by correcting the equations and either using the standard nonnegative Johnston index or explicitly defining a marginal-contribution variant with a different name. Since the reader already assessed the paper as CONDITIONAL and the concern supports that assessment, no change to the verdict is needed. The proposed concrete test—recomputing the standard indexes for the paper's own example—would settle whether the discrepancy is merely typographical or exposes a genuine misimplementation.","tokens_in":14723,"tokens_out":22657,"duration_ms":196966,"concrete_test":"Recompute the Deegan-Packel and Johnston indices for the AF in Figure 10 under complete semantics using the published standard formulas: ρ_i = (1/|M(v)|) Σ_{S∈M_i(v)} 1/|S|, and γ_i = Σ_{S∋i, S winning} 1/κ(S), where κ(S) is the number of critical voters in S. Compare these values with Tables III and IV. If the recomputed Johnston values are nonnegative and differ from the table (e.g., b, d, e should be 0, not −1.16667/−0.5/−0.5), then the tool's 'Johnston' ranking is not the Johnston index and the paper must either correct the implementation or clearly define the variant being used.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim of the paper is a parametric family of ranking-based semantics built on four cooperative-game power indexes. For two of the four indexes, the defining equations in Section II-B are not mathematically well-formed. Equation 3 writes the summation as 'S⊆M_i(v)\\{i}' where M_i(v) is a set of coalitions; S therefore ranges over subsets of a family of coalitions, but v_{Si} and |S| require S to be a coalition of players. The standard Deegan-Packel index is ρ_i(v) = (1/|M(v)|) Σ_{S∈M_i(v)} 1/|S|, which does not involve marginal contributions at all. Equation 4 sums over S⊆N\\{i} and divides by κ(S), the number of critical voters in a coalition that does not contain i; the standard Johnston index credits i in coalitions that contain i, dividing by the number of critical voters in that coalition, and is nonnegative. The manuscript's own implementation notes say the tool computes 'π only with respect to those sets S such that either S or S∪{i} is an extension', which is a marginal-contribution construction appropriate for Shapley/Banzhaf but not for Deegan-Packel or Johnston. This is confirmed by Table IV: under complete semantics, the reported 'Johnston' vI values are negative for b (−1.16667), d (−0.5) and e (−0.5), which cannot arise from the standard Johnston index. Thus the rankings in Tables III and IV do not correspond to the named indexes as conventionally defined, and the property discussion in Section V plus the claims in Theorem 2 and Proposition 1 for π∈{ρ,γ} rest on unspecified variants. The Shapley-value property profile (Theorem 1) is independent of this issue, but the paper's broader parametric-family claim is not.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a ranking-based semantics for abstract argumentation, called PI-based semantics, in which each argument is ranked by the lexicographic comparison of two power-index values computed from characteristic functions vI and vO that mark, respectively, the in-labellings and out-labellings of a selected Dung semantics. The authors claim to instantiate this parametric semantics with four cooperative-game power indexes (Shapley, Banzhaf, Deegan-Packel, Johnston), state an axiom profile for the Shapley instance (Theorem 1), state theorems connecting power-index sign with credulous/sceptical acceptance (Theorem 2 and Proposition 1), describe an implementation inside the ConArg web tool, and illustrate the rankings on a worked example.","tokens_in":15052,"tokens_out":12527,"duration_ms":117407,"significance":"The underlying idea is appealing and potentially useful: it connects ranking-based argumentation with well-studied cooperative-game power indexes and makes classical Dung semantics a parameter of the ranking. The Shapley and Banzhaf columns in Tables I and II are plausible and suggest that the tool works for those indexes. If the formal claims were properly proved and all four indexes were implemented faithfully, the paper would be a useful bridge between computational social choice and abstract argumentation. However, the manuscript as submitted does not support the claims for Deegan-Packel and Johnston: the defining equations are not mathematically well-formed, the reported numerical values are inconsistent with the named indexes, and Theorem 2 is contradicted by the paper's own Table III. No machine-checked proofs, reproducible code, or independent verification is provided for the property theorems.","major_comments":[{"comment":"The Deegan-Packel formula as printed is not well-formed: the summation ranges over S⊆M_i(v)\\setminus{i}, but M_i(v) is a family of coalitions, so the subtraction of the player i from this family is undefined and S would range over subsets of a set of coalitions rather than over coalitions. The standard Deegan-Packel index is ρ_i(v) = (1/|M(v)|) Σ_{S∈M_i(v)} 1/|S|, with no marginal term v_{Si}. The values reported in Table III (for example, the ρ-COM row 0.5, 0, 0.5, 0, 0) are consistent with the standard formula using M(v) = {{a,c}} but do not follow from Eq. (3) as written. The paper must correct Eq. (3), state which definition is actually implemented, and reconcile the implementation note in Section IV-A, where the computation is described in terms of marginal contributions v(S∪{i})−v(S), a construction appropriate for Shapley and Banzhaf but not for Deegan-Packel.","section":"Section II-B, Eq. (3)"},{"comment":"The Johnston index is mis-defined. The standard Johnston index credits player i only in winning coalitions that contain i, dividing by the number of critical voters in that coalition, so it is nonnegative. Equation (4), by contrast, sums marginal contributions over S⊆N\\setminus{i} and divides by κ(S), where κ(S) is the number of critical voters in a coalition that does not contain i. The implementation note in Section IV-A confirms that the tool computes v(S∪{i})−v(S) for all four indexes, and Table IV reports impossible negative Johnston values, e.g., γ-CF v_I(b) = −3.16667 and γ-ADM v_I(b) = −6.16667. Therefore the γ rankings and all γ-related property claims in Sections III and V are not about the Johnston index as conventionally defined. Both the equation and the implementation must be corrected before the parametric-family claim can be assessed.","section":"Section II-B, Eq. (4) and Table IV"},{"comment":"Theorem 2 is false for the Deegan-Packel index under the standard (nonnegative) definition. If an argument is never accepted, it belongs to no minimal winning coalition, so its Deegan-Packel value is 0, not strictly negative as the theorem claims. This is visible in the paper's own Table III: the ρ-COM row gives b the value 0.00000 even though b is never in a complete extension. The statement 'if a is credulously rejected then π_a(v_I) < 0' can hold at best for the Shapley and Banzhaf instances, and the theorem should be re-stated index by index after the Deegan-Packel and Johnston definitions are fixed.","section":"Section III, Theorem 2"},{"comment":"The proof of Proposition 1 is not a valid derivation. From the fact that a credulously accepted argument i has at least one positive marginal contribution, it does not follow that the aggregated index value of i is higher than that of a rejected argument j; the aggregate is a weighted sum over all coalitions and may include many negative terms. A correct proof must compare the full weighted sums, or restrict to indexes for which the comparison can be established. The same gap affects the 'straightforward' proof of Theorem 2. Given that these results are presented as theorems and propositions, the authors need to supply complete proofs from the actual definitions.","section":"Section III, Proposition 1 and its proof"},{"comment":"The proof of Theorem 1 is informal and figure-based, and it is not possible to check the claimed property profile from the text. For example, the Self-contradiction case asserts an inequality involving E and v_I without deriving the Shapley comparison, and the other cases are delegated to Figures 2–6 with no numerical evaluations. Since Theorem 1 is the main theoretical contribution, the authors should provide either formal derivations or a machine-checkable verification script that reproduces each claim. If the paper is intended strictly as a tool report, the property statements should be presented as experimental findings rather than as proven theorems.","section":"Section III, Theorem 1"}],"minor_comments":[{"comment":"The phrase 'well know properties' should be 'well-known properties'.","section":"Abstract and Introduction"},{"comment":"The definition of a strict ranking contains a typo: 'a ≻F b is a shortcut for a ≽F b and b⁄≽F b' should read 'b⁄≽F a'.","section":"Definition 4"},{"comment":"The entries '0, 16667' use a decimal comma instead of a decimal point, which is inconsistent with the other tables and should be corrected.","section":"Table III"},{"comment":"The citation for the Johnston index is imprecise: the text refers to Johnston via reference [21], a general encyclopedia, while the formal definition is attributed to the Durán et al. paper [18]; the citation should be aligned with the actual source of the formula.","section":"Section II-B"}],"recommendation":"major_revision","confidential_remarks":"This is a preliminary tool report, but it makes substantive theoretical claims. The Shapley-based part is plausible and the tool availability is a positive feature. The Deegan-Packel and Johnston parts are in their current form not just under-proved: they are mis-defined, and the paper's own Table III contradicts Theorem 2. These issues are fixable in principle by correcting the definitions, re-running the tool, and re-checking the theorems, so I recommend major revision rather than rejection. The authors should also be asked to provide reproducible verification for Theorem 1."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two quick things. The central idea—rank arguments by Shapley/Banzhaf/Deegan-Packel/Johnston power indexes computed over the extensions of a chosen Dung semantics—is a natural and genuinely useful transfer. The paper is the first place I've seen it worked out beyond the authors' own Shapley-only CEUR sketches. The ConArg tool with the greyscale output is a nice practical addition, and the property analysis in Theorem 1, informal as it is, starts to map the space properly.\n\nThe problem is that the paper overreaches as written. Equations 3 and 4 are not the Deegan-Packel and Johnston indexes. Equation 3 sums over subsets of a family of coalitions and uses marginal contributions where the standard DP index uses none; Equation 4 divides by κ(S) for coalitions that don't contain i, which inverts the standard construction. The implementation note in Section IV-A says the tool computes 'π only with respect to those sets S such that either S or S∪{i} is an extension', which is the right shortcut for Shapley and Banzhaf but not for DP and Johnston. The negative 'Johnston' values in Table IV confirm something is off: the standard Johnston index is nonnegative. So the rankings in Tables III and IV for ρ and γ are not the rankings promised by the names. That undercuts the parametric-family claim, though the Shapley-based results stand.\n\nThe other soft spot is the acceptance theorems. Proposition 1's proof shows that a credulously accepted argument has at least one positive marginal contribution; it does not show its total index is higher than that of a rejected argument. The proof of Theorem 2 is just declared 'straightforward'. These claims may be true, but they are unproven in this text. Theorem 1's property proofs lean on figures; for a preliminary report that is tolerable, but a full paper would need real case analysis or machine-checked proofs.\n\nThe citation pattern is fine—the authors build on their own prior work and credit the relevant game-theory and argumentation literature. Nothing circular.\n\nBottom line: this is a preliminary report with a good idea and a flawed execution. It deserves a serious referee, and I would send it to review, but with a request for major revision: fix the DP and Johnston formulas and re-run the tables, and prove or refute Theorems 2 and Proposition 1. If the authors do that, this becomes a solid contribution to ranking-based semantics. I wouldn't cite it as it stands; I'd point students to it as an interesting problem statement.","headline":"A useful but overreaching preliminary report: the Shapley-based ranking is a good idea, but the Deegan-Packel and Johnston formulas are wrong as printed and the acceptance theorems lack valid proofs.","tokens_in":15640,"tokens_out":3294,"would_cite":false,"duration_ms":113935,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper defines a ranking-based semantics that turns Dung labellings into cooperative games and ranks arguments by one of four power indexes, proving the property profile for the Shapley case.","keywords":["ranking-based semantics","abstract argumentation","power indexes","Shapley value","Banzhaf index","Deegan-Packel index","Johnston index","Dung semantics"],"falsifier":"For the AF in Figure 10, compute the Deegan-Packel and Johnston values from the standard definitions (minimal winning coalitions for $\\rho$; sum of $1/\\kappa(S)$ over winning coalitions for each critical voter for $\\gamma$) using the listed extensions, and compare with Tables III and IV: any mismatch shows the implemented indexes are not those named. For Theorem 1, a finite AF giving a violation of any listed property row would refute that row.","tokens_in":14455,"feed_emoji":"🗳️","tokens_out":11293,"duration_ms":290218,"temperature":0.7,"pith_summary":"Ranking-based semantics in abstract argumentation usually sort arguments by graph-theoretic criteria with no tie to classical Dung acceptability. This paper claims that the classical semantics themselves can drive the ranking: for any chosen Dung semantics $\\sigma$ and any power index $\\pi$, it builds a two-sided cooperative game over the arguments, ranking each argument by $\\pi$'s value on the 'in' extensions and breaking ties by the value on the 'out' extensions. The resulting PI-based semantics is total for the Shapley value and satisfies Abstraction, Independence and Totality for five standard Dung semantics; it satisfies Self-contradiction only for conflict-free and Non-attacked Equivalence only for complete, preferred, and stable semantics, and never satisfies Cardinality Precedence or Quality Precedence. The paper also reports an implementation of all four indexes in the ConArg web tool and shows the rankings on a worked example. If correct, this gives a parameterised family of argument rankings that inherit meaning from classical acceptance criteria, useful when a large framework must be reduced to a shortlist.","feed_headline":"Four voting-power indexes now rank arguments in Dung frameworks","feed_subtitle":"ConArg implements Shapley, Banzhaf, Deegan-Packel, and Johnston rankings tied to classical Dung semantics.","key_machinery":"The engine is the pair of characteristic functions $v^I_{\\sigma,F}$ and $v^O_{\\sigma,F}$ (Definition 9) together with the lexicographic comparison in Definition 10. Instead of introducing a new graph heuristic, the semantics turns the set of extensions of a chosen Dung semantics into a simple voting game, then asks how often and how decisively each argument belongs to winning coalitions. The Shapley value weights marginal contributions by the number of orderings in which a player joins a coalition; Banzhaf counts critical voters without ordering; Deegan-Packel averages over minimal winning coalitions; Johnston gives critical voters a fractional $1/\\kappa(S)$ share. The implementation computes $\\pi$ only for coalitions $S$ where either $S$ or $S\\cup\\{i\\}$ is an extension, which is exactly where the marginal contribution can be non-zero.","core_discovery":"The central claim is that a ranking-based semantics can be defined by taking the in/out labellings of a Dung semantics $\\sigma$ as the winning coalitions of a simple game over the arguments: $v^I_{\\sigma,F}(S)=1$ exactly when $S$ is an in-set of some $\\sigma$-labelling, and $v^O_{\\sigma,F}(S)=1$ exactly when $S$ is an out-set of some $\\sigma$-labelling. A power index $\\pi$ (Shapley, Banzhaf, Deegan-Packel, or Johnston) is then evaluated on these two characteristic functions, and two arguments are compared lexicographically: higher $\\pi(v^I_{\\sigma,F})$ is better, and in case of a tie, lower $\\pi(v^O_{\\sigma,F})$ is better. The paper proves, for the Shapley value, that Abstraction, Independence and Totality hold for conflict-free, admissible, complete, preferred and stable semantics; Self-contradiction holds only for conflict-free; Non-attacked Equivalence holds only for complete, preferred and stable; and neither Cardinality Precedence nor Quality Precedence ever holds. It further claims that sceptically accepted arguments receive positive power-index value, credulously rejected arguments receive negative value, and consequently the semantics satisfies $\\delta$-Sceptical Precedence and $\\delta$-Credulous Precedence for every $\\delta$ among the five Dung semantics.","pith_inferences":["Because $v^I$ and $v^O$ are defined only on exact extension sets, an argument's power-index value is determined entirely by the collection of extensions, not by graph paths; a natural testable extension is to weight coalitions by defence chains or attack distances and see whether Cardinality or Quality Precedence begin to hold.","Only the Shapley case receives a theorem in this paper; the Banzhaf, Deegan-Packel and Johnston rankings are demonstrated through tables, so a full property profile for those three indexes remains open rather than established.","The 'out' tie-breaker rewards arguments that are defeated in fewer ways, not just accepted more often, so the semantics can distinguish arguments that look identical when only in-extensions are compared; this could be exploited in frameworks with multiple preferred extensions.","The implementation's pruning of coalitions to those that are extensions may avoid the generic NP-hardness of power-index computation, and combining the Shapley value with restricted-coalition game values would let the ranking ignore arguments that are not even credulously accepted."],"forward_implications":["For any framework and any of the five Dung semantics, the Shapley-based semantics returns a total ordering of arguments, so no argument is left incomparable.","Under complete, preferred and stable semantics, all non-attacked arguments receive the same rank; under conflict-free semantics, self-attacking arguments are strictly worse than non-self-attacking ones.","More numerous or higher-ranked direct attackers do not by themselves make an argument worse: Cardinality Precedence and Quality Precedence fail because an argument's value includes how many other acceptable arguments it defends.","Sceptically accepted arguments are ranked above merely credulous or rejected arguments, and credulously accepted arguments above rejected ones, for all four power indexes and all five Dung semantics.","Unlike the Cat, Dbs and Bds ranking semantics, the PI-based semantics satisfies Credulous Precedence for admissible, complete, preferred and stable semantics."],"supporting_citations":[{"why":"It supplies the ranking-property framework (Abstraction, Independence, Self-contradiction, Cardinality Precedence, Quality Precedence, Non-attacked Equivalence, Totality) used to evaluate the PI-based semantics.","marker":"[2]"},{"why":"It defines reinstatement labelling, the source of the in/out sets that the characteristic functions are built on.","marker":"[12]"},{"why":"It defines abstract argumentation frameworks and the Dung extension semantics that parameterise the ranking.","marker":"[16]"},{"why":"It defines the Shapley value, the power index for which Theorem 1's property profile is proved.","marker":"[27]"},{"why":"It defines the Banzhaf index used for the $\\beta$ rankings.","marker":"[3]"},{"why":"It provides the interpretation of minimal winning coalitions used in the Deegan-Packel formula to avoid division by zero.","marker":"[1]"},{"why":"It is the cited source for the Deegan-Packel and Johnston power indexes implemented in the tool.","marker":"[21]"},{"why":"It supplies the Johnston-index computation based on critical-voter counts used in Equation 4.","marker":"[18]"},{"why":"It is the ConArg constraint-based solver that returns the extensions (winning coalitions) consumed by the ranking script.","marker":"[10]"}],"fun_headline_variants":["ConArg ranks Dung arguments with four power indexes","Power indexes now sort arguments in ConArg's ranking semantics","Shapley, Banzhaf, Deegan-Packel, Johnston drive ConArg ranking","Voting power indexes give arguments a ranking in ConArg","ConArg's new ranking: power indexes decide argument order"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the Deegan-Packel and Johnston power indexes are faithfully transcribed and implemented (Equations 3 and 4), yet those formulas as written are not well-formed: Equation 3 sums over a set-of-coalitions expression that does not parse, and Equation 4 divides by the critical-voter count of coalitions that do not contain the player, so a mistranscription would make the reported $\\rho$ and $\\gamma$ rankings something other than the named indexes.","fun_headline_variants_meta":{"raw":{"variants":["ConArg ranks Dung arguments with four power indexes","Power indexes now sort arguments in ConArg's ranking semantics","Shapley, Banzhaf, Deegan-Packel, Johnston drive ConArg ranking","Voting power indexes give arguments a ranking in ConArg","ConArg's new ranking: power indexes decide argument order"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000896,"raw_usage":{"total_tokens":3859,"prompt_tokens":939,"completion_tokens":2920,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":555,"completion_tokens_details":{"reasoning_tokens":2831}},"tokens_in":555,"tokens_out":2920,"duration_ms":20928,"temperature":1.0,"reasoning_tokens":2831,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:57:28.320169+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"For the AF in Figure 10, compute the Deegan-Packel and Johnston values from the standard definitions (minimal winning coalitions for $\\rho$; sum of $1/\\kappa(S)$ over winning coalitions for each critical voter for $\\gamma$) using the listed extensions, and compare with Tables III and IV: any mismatch shows the implemented indexes are not those named. For Theorem 1, a finite AF giving a violation of any listed property row would refute that row.","supporting_citations":[{"cited_title":"Amgoud and J","cited_arxiv_id":null,"evidence_quote":"It supplies the ranking-property framework (Abstraction, Independence, Self-contradiction, Cardinality Precedence, Quality Precedence, Non-attacked Equivalence, Totality) used to evaluate the PI-based semantics."},{"cited_title":"Caminada","cited_arxiv_id":null,"evidence_quote":"It defines reinstatement labelling, the source of the in/out sets that the characteristic functions are built on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It defines abstract argumentation frameworks and the Dung extension semantics that parameterise the ranking."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It defines the Shapley value, the power index for which Theorem 1's property profile is proved."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It defines the Banzhaf index used for the $\\beta$ rankings."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It provides the interpretation of minimal winning coalitions used in the Deegan-Packel formula to avoid division by zero."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It is the cited source for the Deegan-Packel and Johnston power indexes implemented in the tool."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It supplies the Johnston-index computation based on critical-voter counts used in Equation 4."},{"cited_title":"Bistarelli and F","cited_arxiv_id":null,"evidence_quote":"It is the ConArg constraint-based solver that returns the extensions (winning coalitions) consumed by the ranking script."}],"review_version":1}