Argument Defensibility

• Argument defensibility is defined as measuring an argument’s ability to withstand critical scrutiny within structured frameworks like Dung's abstract argumentation.
• Graded and ranking-based semantics offer fine-grained analysis by quantifying internal conflicts and assigning real-valued defensibility scores.
• Practical applications include dialogue-based defenses, requirements engineering, and automated judgment systems ensuring transparent and policy-compliant argument evaluation.

Argument defensibility is the formal and practical characterization of how robustly an argument can withstand attack, defeat, or critical scrutiny within a structured argumentation framework. It encompasses both the qualitative acceptability of an argument—whether its chain of reasoning can be maintained under all permissible challenges—and the quantitative degree or ranking of its justification. Modern research offers a richly stratified landscape of approaches, ranging from extension-based and defense-aware semantics in abstract argumentation to numerical ranking and dialogue-based tenability, as well as operationalizations in applied domains such as requirements engineering, compliance, and automated judgment.

1. Fundamental Definitions and Frameworks

At its core, argument defensibility is shaped by the formalism of abstract argumentation frameworks (AFs) in the sense of Dung, where a set of arguments $A$ and a binary attack relation $R \subseteq A \times A$ provide the basis for determining which arguments are "in" under various semantics (grounded, preferred, stable, etc.). Defensibility is traditionally linked to two notions:

• Conflict-freeness: An argument set $S$ is conflict-free iff no $x, y \in S$ with $(x, y) \in R$.
• Defense: $S$ defends $a$ if for every attacker $b$ of $a$, there exists $c \in S$ attacking $R \subseteq A \times A$0.

Under Dung-style semantics, an argument is defensible if it lies in at least one admissible extension; that is, it can consistently be included in a conflict-free set that rebuts every attack using the available counter-arguments (Grossi et al., 2018, Amgoud et al., 2013). However, this uniform defense requirement has recognized limitations in scenarios where strategic, non-uniform defense is more realistic—e.g., when different attackers are mutually inconsistent, and the proponent could respond differently depending on which challenge is posed (Andrews et al., 3 May 2026).

Graded variants further generalize these definitions: for integers $R \subseteq A \times A$1, one can define $R \subseteq A \times A$2-conflict-free and graded $R \subseteq A \times A$3-defense, permitting more fine-grained thresholds for allowable internal attacks and the minimum number of defenders for each attack, leading to a spectrum of defensibility standards (Grossi et al., 2018).

2. Graded and Ranking-Based Semantics

To overcome the binary nature of classical extension-based acceptability, graded semantics introduce fine-grained controls and rankings:

• Graded Conflict-Freeness: $R \subseteq A \times A$4 is $R \subseteq A \times A$5-conflict-free iff $R \subseteq A \times A$6, permitting up to $R \subseteq A \times A$7 internal attacks.
• Graded Defence: $R \subseteq A \times A$8 collects arguments for which fewer than $R \subseteq A \times A$9 attackers are defended by fewer than $S$0 defenders in $S$1.
• Graded Extensions: Sets meeting both $S$2-conflict-freeness and $S$3-defense constitute $S$4-admissible sets; those that are also fixed points are $S$5-complete.

This approach yields not only a variety of extension types (e.g., $S$6-grounded, preferred) but also defensibility rankings. Contextual and absolute argument rankings reflect how robustly an argument would retain acceptability as standards are tightened:

• $S$7 assigns a real-valued defensibility score, indicating the maximal number of defenders supporting $S$8 under increasingly strict thresholds (Grossi et al., 2018).

Ranking-based semantics generalize this to real-valued or ordinal measures, supporting comparative assessments of argument strength even when extension sets are not unique or overlap (Grossi et al., 2018, Pu et al., 2015). The attacker–defender counting approach, for example, propagates influence via walks in the attack graph (signed and damped), converging to a unique vector of defensibility scores across all arguments (Pu et al., 2015).

3. Defense Structure, Justification, and Reason Tracing

Moving beyond extension membership, defense semantics and their variants focus on explicit justifications—detailing not only that but also why an argument is accepted:

• Defense Graphs: Nodes represent defense claims ("$S$9 defends $x, y \in S$0 by attacking $x, y \in S$1"), and attacks among defenses reflect failure propagation across chains of attack and defense (Liao et al., 2017, Liao et al., 2023).
• Direct and Root Reasons: For each argument, one can compute the sets of direct defenders (immediate reasons for acceptance) and root defenders (ultimate undefeated supporters), thereby mapping the defense tree or chain for every argument in every defense extension (Liao et al., 2017, Liao et al., 2023).

This fine-grained structuring enables new equivalence notions:

• Defense Equivalence: Two AFs are equivalent if they generate the same sets of defense extensions, which is strictly finer than standard or strong equivalence under classical Dung semantics.
• Root Equivalence: Two AFs are root-equivalent on a subset of arguments if for each argument, the set of ultimate defenders is the same, facilitating argument summarization and modularity (Liao et al., 2023).

Defense-justification frameworks have direct application in requirements engineering and regulated environments, supporting traceability and auditability from final decisions back through the exact sequence of attacks and defenses that justified acceptance (Cheng et al., 25 Apr 2026).

4. Dialogue, Tenability, and Strategic Defensibility

Recent work emphasizes non-uniform, dialogue-based approaches to defensibility—capturing situations where a proponent maintains an argument against every individual conflict-free attack, rather than demanding a single uniform defense set (Andrews et al., 3 May 2026).

• Tenability Semantics: Static tenability formalizes that for every finite conflict-free challenger set $x, y \in S$2, there exists a conflict-free superset countering each attack in $x, y \in S$3; strong tenability extends this to dialogue games (commitment games with monotonicity). This relaxes classical admissibility, admitting arguments that can always be defended via some winning (even non-uniform) strategy.
• Critical Benchmarks: Tenability distinguishes itself by correctly handling classic anomalies such as self-defeating attacks, floating assignment, and disjunctive reinstatement, where uniform extension-based semantics are excessively strict (Andrews et al., 3 May 2026).

These dialogue-sensitive notions yield a spectrum of defensibility—e.g., from admissibility (strong), through tenability (weaker), down to static tenability (weakest). Complexity results highlight the computational boundaries of evaluating these properties.

5. Practical and Applied Metrics for Defensibility

Operationalizing defensibility requires metrics and methodologies adapted to context:

• Policy-Grounded Defensibility: In rule-governed AI/ML settings, defensibility is the proportion of decisions defensible under explicit, checkable derivations from the written rulebook (Defensibility Index, DI). Complementary metrics include the Ambiguity Index (AI) and the Probabilistic Defensibility Signal (PDS), which decompose defensibility into policy compliance, intrinsic rule ambiguity, and per-case reasoning stability (O'Herlihy et al., 22 Apr 2026).
• Defeaters and Assurance: Assurance frameworks employ explicit defeater nodes, which must be resolved or refuted for claims to remain supported. Multi-level defeater handling and eliminative argumentation systematically capture, address, and trace every doubt or possible objection, algorithmically enforcing defensibility of assurance cases (Bloomfield et al., 2024).
• Argument Vulnerability: Machine learning techniques quantify the "attackability" of argument components, mapping linguistic, semantic, and pragmatic features to probabilistic attackability and thereby suggesting which arguments or sentences are most (or least) defensible (Jo et al., 2020).

In dialogue systems and multi-agent negotiation (e.g., ArgRE), defended arguments correspond to those in the accepted extension under grounded or preferred semantics, with the AF's structure influencing whether skepticism (grounded) or credulity (preferred) governs the defensibility standard (Cheng et al., 25 Apr 2026).

6. Argument Ranking, Aggregation, and Structural Robustness

Defensibility scores are also computed via propagation and fixed-point procedures over explicit interaction graphs:

• GRASP Framework: Aggregates local attack/defense judgments (elicited via LLMs or human annotators) into a global, convergent ranking by iterative propagation of support and attack weights. Structural sufficiency is defined as the property that every attacker of an argument is itself attacked; GRASP generalizes this via continuous scores captured at the fixed point of the propagation operator (Misra et al., 18 May 2026).
• Axiomatic Characterization: Structural sufficiency satisfies attack sensitivity, defense reinstatement, and structural locality, and is strictly divorced from persuasion or factuality, providing a measure of an argument's defense-aware robustness in the explicit graph.

This approach ensures that defensibility is not conflated with rhetorical effect, enabling transparent, reproducible, and auditably consistent argument ranking for both human and automated judges.

7. Interpretative and Rule-Based Defensibility

In the domain of normative rule-following—especially in law and governance—interpretive argumentation has developed the notion of minimally defeasible interpretive arguments (MDIA) (Licato, 2021). Here, a candidate interpretation $x, y \in S$4 of an open-textured rule is maximally defensible if every potential defeater for $x, y \in S$5 is itself defeated by at least one support argument, and the set of supports is minimal. Residual defeasibility and aggregated support scores are formally compared across interpretations to select the maximally justified reading, ensuring that the AI or system acts under the most defensibly grounded interpretation available.

• (Grossi et al., 2018) On the Graded Acceptability of Arguments in Abstract and Instantiated Argumentation
• (Amgoud et al., 2013) On the Acceptability of Arguments in Preference-Based Argumentation
• (Cheng et al., 25 Apr 2026) ArgRE: Formal Argumentation for Conflict Resolution in Multi-Agent Requirements Negotiation
• (O'Herlihy et al., 22 Apr 2026) Escaping the Agreement Trap: Defensibility Signals for Evaluating Rule-Governed AI
• (Roos, 2022) On resolving conflicts between arguments
• (Na et al., 2022) The Diversity of Argument-Making in the Wild
• (Liao et al., 2017) Defense semantics of argumentation: encoding reasons for accepting arguments
• (Liao et al., 2023) Defense semantics of argumentation: revisit
• (Pu et al., 2015) Attacker and Defender Counting Approach for Abstract Argumentation
• (Andrews et al., 3 May 2026) Tenability and Weak Semantics: Modeling Non-uniform Defense
• (Jo et al., 2020) Detecting Attackable Sentences in Arguments
• (Misra et al., 18 May 2026) GRASP: Deterministic argument ranking in interaction graphs
• (Bloomfield et al., 2024) Defeaters and Eliminative Argumentation in Assurance 2.0
• (Licato, 2021) How Should AI Interpret Rules? A Defense of Minimally Defeasible Interpretive Argumentation