Contradiction Index: Quantifying Inconsistency

• Contradiction Index is a family of quantitative constructs that measure internal inconsistency and incompatibility across various formal objects using precise mathematical and operational approaches.
• It employs methodologies such as contradiction graphs, belief-function distances, and evidence aggregation to capture graded indices, including an exact equivalence with VC dimension in some settings.
• Applications span diverse domains from statistical learning and NLP to peer-review analysis, facilitating calibrated and evidence-grounded contradiction detection for informed decision-making.

Contradiction Index denotes a family of quantitative constructs for measuring incompatibility, internal inconsistency, or contradiction within a formal object. The object may be a concept class, a basic belief assignment, an article, a dialogue, a multimodal input, an enterprise document collection, or a pair of scientific reviews. The literature does not supply a single domain-independent definition. Instead, it provides several mathematically precise indices and many derived operationalizations: in statistical learning theory, a contradiction-based index induced from contradiction graphs coincides exactly with VC dimension; in belief-function theory, a contradiction measure quantifies self-opposition of a mass function; and in NLP, multimodal reasoning, and evaluation benchmarks, contradiction indices are typically scalar summaries obtained by aggregating contradiction probabilities, evidence spans, or intensity labels (Campbell et al., 19 May 2026, Smarandache et al., 2011, Kumar et al., 11 May 2026).

1. Terminological scope and general form

The term is used heterogeneously across fields. Some works define a contradiction-based quantity explicitly or canonically. Others do not use the term “Contradiction Index” in the paper text, but their formalism induces one directly. This is explicit in the contradiction-graph treatment of concept classes, where the paper states that it does not use the term “Contradiction Index” explicitly but canonically induces the index

$$\mathrm{CI}(H):=\sup\left\{m\in\mathbb{N}_0:\ \mathcal{P}_m\bigl(G_m(H)\bigr)\right\},$$

and proves that it equals VC dimension. A similar pattern appears in several NLP works, where the published model outputs contradiction probabilities or intensity labels and a scalar CI is then defined as an aggregation of those outputs rather than as a native training objective (Campbell et al., 19 May 2026, Hsu et al., 2021, Lendvai et al., 2016).

Across the cited literature, contradiction indices serve three recurrent functions. First, they act as threshold certificates, as in the graph-theoretic case where the presence of a specific graph pattern certifies whether $\mathrm{VCdim}(H)\ge m$. Second, they act as normalized scalar summaries, such as weighted averages of contradiction probabilities, contradiction rates, or distances to categorical reference points. Third, they act as graded evidence aggregators, especially in review analysis and multimodal evaluation, where contradiction is localized, typed, and assigned ordinal intensity before aggregation. This suggests that “Contradiction Index” is best understood as a design pattern for contradiction quantification rather than a single invariant.

2. Contradiction graphs and the exact recovery of VC dimension

For a binary concept class $H \subseteq \{0,1\}^X$, the order-$m$ contradiction graph $G_m(H)$ has as vertices the $H$-realizable labeled sequences of length $m$, and two vertices are adjacent when they assign opposite labels to some shared domain point. If

$$\mathcal{P}_m(G)\quad\text{means “G contains a cube-trace clique of size }2^m\text{,”}$$

then the central theorem is

$$\mathrm{VCdim}(H)\ge m\quad\iff\quad \mathcal{P}_m\bigl(G_m(H)\bigr).$$

The witness is not merely a large clique. It is a clique $Q$ of size $\mathrm{VCdim}(H)\ge m$0 together with a bijection $\mathrm{VCdim}(H)\ge m$1 such that, for every vertex $\mathrm{VCdim}(H)\ge m$2, the non-neighbor trace $\mathrm{VCdim}(H)\ge m$3 maps under $\mathrm{VCdim}(H)\ge m$4 to a Boolean subcube. The forward direction uses shattered sets to construct such cliques; the reverse direction uses repeated vertices $\mathrm{VCdim}(H)\ge m$5 and $\mathrm{VCdim}(H)\ge m$6, complementary-facet rigidity of Boolean subcubes, and a counting argument to force a common support of exactly $\mathrm{VCdim}(H)\ge m$7 informative points (Campbell et al., 19 May 2026).

This yields a contradiction-based index

$\mathrm{VCdim}(H)\ge m$8

with levelwise indicator

$\mathrm{VCdim}(H)\ge m$9

The resulting identity

$H \subseteq \{0,1\}^X$0

means that contradiction graphs determine the full VC dimension exactly, including the finite-versus-infinite dichotomy. If $H \subseteq \{0,1\}^X$1 for all $H \subseteq \{0,1\}^X$2, then $H \subseteq \{0,1\}^X$3. This answers the question of Alon–Moran–Schefler–Yehudayoff about whether contradiction graphs at all levels could fail to distinguish finite from infinite VC dimension.

Several examples delimit the construction. Threshold functions on $H \subseteq \{0,1\}^X$4 satisfy $H \subseteq \{0,1\}^X$5, so $H \subseteq \{0,1\}^X$6 holds and $H \subseteq \{0,1\}^X$7 fails. Intervals on $H \subseteq \{0,1\}^X$8 satisfy $H \subseteq \{0,1\}^X$9, so $m$0 holds and $m$1 fails. Axis-aligned rectangles in $m$2 satisfy $m$3, and halfspaces in $m$4 satisfy $m$5. The parity example shows that a finite prefix of contradiction graphs need not determine exact VC dimension: for parity-related classes $m$6 and $m$7, one can have $m$8 for all $m$9, even though $G_m(H)$0 and $G_m(H)$1. Conversely, the “left-turn indicators” tree class has $G_m(H)$2 but $G_m(H)$3 contains $G_m(H)$4-cliques for all $G_m(H)$5; these fail the cube-trace condition, showing that clique number alone is insufficient.

3. Contradiction as self-opposition in belief-function theory

In the theory of belief functions, Smarandache, Martin, and Osswald introduce a contradiction measure for a single basic belief assignment $G_m(H)$6, explicitly distinguishing it from conflict between multiple bbas. For any focal element $G_m(H)$7, let $G_m(H)$8 be the categorical bba at $G_m(H)$9. Using the Jousselme distance

$H$0

with

$H$1

the contradiction of $H$2 with respect to $H$3 is

$H$4

The contradiction of the bba itself is then

$H$5

where $H$6 is a normalization constant chosen to map the index to $H$7; in the paper’s illustrations using Jousselme distance on $H$8, $H$9 (Smarandache et al., 2011).

This measure is an index of internal inconsistency. It is zero for any categorical bba, since the distance to its own categorical representative vanishes. It is also zero for total ignorance $m$0, because this too is categorical on $m$1. By contrast, balanced mass on mutually exclusive singletons yields high contradiction: for $m$2 and $m$3, the paper obtains $m$4 under the stated normalization. The measure decreases as mass moves from precise focal sets to more imprecise sets, because the Jaccard-overlap structure of $m$5 shrinks distances when focal elements share larger intersections.

The same paper places contradiction alongside companion uncertainty measures. The normalized degree of non-specificity is

$m$6

the degree of Bayesianity is

$m$7

and the degree of specificity is defined by distance to a “most specific associated mass” $m$8,

$m$9

These measures separate different notions often conflated under “uncertainty”: contradiction quantifies self-opposition, non-specificity quantifies imprecision, Bayesianity quantifies concentration on singletons, and specificity quantifies proximity to a categorical representative. The paper further uses $$\mathcal{P}_m(G)\quad\text{means “G contains a cube-trace clique of size }2^m\text{,”}$$0 to characterize fusion rules.

4. Pairwise, article-level, and conversational operationalizations

In document-level contradiction detection, contradiction indices are commonly derived from model outputs. In “WikiContradiction,” an article is represented as a sequence of sentences $$\mathcal{P}_m(G)\quad\text{means “G contains a cube-trace clique of size }2^m\text{,”}$$1, a pairwise contradiction module outputs probabilities

$$\mathcal{P}_m(G)\quad\text{means “G contains a cube-trace clique of size }2^m\text{,”}$$2

Top-$$\mathcal{P}_m(G)\quad\text{means “G contains a cube-trace clique of size }2^m\text{,”}$$3 pairs are selected, and a self-attention layer aggregates them to an article representation. The paper synthesis defines several article-level indices from these components, including a mean-based index, an attention-weighted index, a length-normalized index, and a max-based index:

$$\mathcal{P}_m(G)\quad\text{means “G contains a cube-trace clique of size }2^m\text{,”}$$4

together with $$\mathcal{P}_m(G)\quad\text{means “G contains a cube-trace clique of size }2^m\text{,”}$$5 and $$\mathcal{P}_m(G)\quad\text{means “G contains a cube-trace clique of size }2^m\text{,”}$$6. Here the CI is a transparent proxy for article-level self-contradiction, and the Top-$$\mathcal{P}_m(G)\quad\text{means “G contains a cube-trace clique of size }2^m\text{,”}$$7 list supplies localized evidence (Hsu et al., 2021).

For rumor-related Twitter data, contradiction is cast as a 3-way RTE task with labels Entailment, Contradiction, and Unknown. The pairwise contradiction index is simply

$$\mathcal{P}_m(G)\quad\text{means “G contains a cube-trace clique of size }2^m\text{,”}$$8

and set-level or thread-level indices are aggregated over relevant tweet pairs. In the threaded case, the synthesis defines

$$\mathcal{P}_m(G)\quad\text{means “G contains a cube-trace clique of size }2^m\text{,”}$$9

so that contradiction directed at the source claim and contradiction internal to the thread can be combined with structural weights (Lendvai et al., 2016).

The same aggregation logic appears in dialogue generation. In the analysis of contradiction-awareness in neural response generation, the published metrics are Certainty

$$\mathrm{VCdim}(H)\ge m\quad\iff\quad \mathcal{P}_m\bigl(G_m(H)\bigr).$$0

and Variety

$$\mathrm{VCdim}(H)\ge m\quad\iff\quad \mathcal{P}_m\bigl(G_m(H)\bigr).$$1

where $$\mathrm{VCdim}(H)\ge m\quad\iff\quad \mathcal{P}_m\bigl(G_m(H)\bigr).$$2 contains inputs whose $$\mathrm{VCdim}(H)\ge m\quad\iff\quad \mathcal{P}_m\bigl(G_m(H)\bigr).$$3-best lists include at least one noncontradictory response. A derived single-number contradiction-awareness index is

$$\mathrm{VCdim}(H)\ge m\quad\iff\quad \mathcal{P}_m\bigl(G_m(H)\bigr).$$4

equivalently the average noncontradictory fraction across all inputs. This is not named in the original paper, but it is directly induced by the published definitions (Sato et al., 2022).

Low-resource NLI work treats contradiction indices similarly. In Arabic NLI, the paper does not define a contradiction index, but its architecture naturally supports the derived quantity

$$\mathrm{VCdim}(H)\ge m\quad\iff\quad \mathcal{P}_m\bigl(G_m(H)\bigr).$$5

with optional logit, margin, or $$\mathrm{VCdim}(H)\ge m\quad\iff\quad \mathcal{P}_m\bigl(G_m(H)\bigr).$$6-score variants. The model combines a contradiction feature vector—Arabic named-entity features, synonym/neutral/antonym counts, special stopword features, and number/date/time features—with a language-model vector before classification (Jallad et al., 2022). In Persian contradiction detection, the paper likewise does not define a CI, but it supplies category-aware detectors over Negation, Numeric, Antonym, Structural, and Others, making a contradiction rate or category-wise contradiction proportion the natural aggregation (Rahimi et al., 2021). In Chinese conversations, CDConv does not define a named index either, but the annotation protocol over Intra-sentence Contradiction, Role Confusion, and History Contradiction directly supports a contradiction rate over bot turns or conversations; on the released benchmark, contradictions account for 7,309 of 11,660 conversations (Zheng et al., 2022).

A structurally different textual formalization appears in the sheaf model of contradictions and disagreements. There, contradiction is tied to the non-existence of global sections over a poset of theories with shared predicates. The minimal binary index is

$$\mathrm{VCdim}(H)\ge m\quad\iff\quad \mathcal{P}_m\bigl(G_m(H)\bigr).$$7

and a graded partial contradiction measure is

$$\mathrm{VCdim}(H)\ge m\quad\iff\quad \mathcal{P}_m\bigl(G_m(H)\bigr).$$8

where $$\mathrm{VCdim}(H)\ge m\quad\iff\quad \mathcal{P}_m\bigl(G_m(H)\bigr).$$9 is the size of the largest subset of theories admitting a global section. This makes contradiction a failure of gluing and disagreement a matter of local versus global reconcilability (Zadrozny et al., 2018).

5. Multimodal, enterprise, and peer-review indices

In multimodal contradiction detection, the CLASH benchmark does not define a named contradiction-specific metric, but it motivates one from per-category conflict-detection recall. The proposed base index is a macro-average over object and attribute categories,

$Q$0

with micro-averaged and penalized variants. The recommended composite form is

$Q$1

where $Q$2 is a modality-bias summary over failures. Here the contradiction index is not merely detection accuracy; it incorporates category coverage, false alarms, format compliance, and failure-mode asymmetry between image-grounded and text-grounded errors (Popordanoska et al., 24 Nov 2025).

In enterprise RAG evaluation, ContraGen provides an explicit pair-level contradiction score. A candidate sentence pair receives an NLI judgment and an LLM-judge judgment, combined by

$Q$3

with

$Q$4

and threshold $Q$5. The paper itself does not define a single named contradiction index, but it explicitly enables document-level self-contradiction rates, cross-document contradiction rates, type-weighted rates over Temporal, Numerical, Authority, Process, Policy Reversal, and Specificity, and retrieval-verifiability-stratified rates (Mantravadi et al., 3 Oct 2025).

The most elaborate graded formulation appears in scientific peer review. RevCI annotates review pairs with contradiction evidence spans, six aspects—Motivation, Clarity, Soundness, Substance, Originality, and Meaningful Comparison—and ordinal intensities $Q$6, $Q$7, and $Q$8, with $Q$9 used internally for invalid contradiction candidates. IMPACT extracts aspect-conditioned evidence, runs deliberative intensity scoring and adjudication, and retains only valid contradictions after a contradiction validity gate. The paper then derives a review-pair contradiction index as a weighted mean of normalized intensities,

$\mathrm{VCdim}(H)\ge m$00

with per-aspect variants

$\mathrm{VCdim}(H)\ge m$01

This makes contradiction an evidence-grounded, aspect-localized, severity-aware object rather than a binary label on isolated sentence pairs (Kumar et al., 11 May 2026).

6. Design principles, robustness, and unresolved issues

A recurrent theme is that simpler surrogates are often insufficient. In contradiction graphs, large cliques alone do not certify shattering; the cube-trace condition is essential. In dialogue contradiction, sentence-pair modeling underuses context, while flattened context adds noise; hierarchical context aligned to contradiction type performs better. In peer-review analysis, isolated sentence-pair contradiction misses review-level aspect structure and graded severity. These results jointly indicate that contradiction indices are most informative when they preserve the relational structure actually responsible for incompatibility (Campbell et al., 19 May 2026, Zheng et al., 2022, Kumar et al., 11 May 2026).

Another theme is calibration versus artifact exploitation. In SNLI adversarial analysis, the contradiction class exhibits a smaller accuracy decline under universal adversarial attacks than entailment and neutral. The paper details derived robustness indices such as

$\mathrm{VCdim}(H)\ge m$02

and shows that contradiction’s apparent resilience is tied in part to lexical artifacts such as negation and antonymy cues rather than to unambiguous semantic depth (Verma et al., 2024). This complicates interpretation: a high contradiction index may reflect robust contradiction modeling, but it may also reflect robust exploitation of spurious correlations.

Several works therefore emphasize typology, evidence, and human oversight. Prototype generation for contradiction detection uses rule-based negation, antonymy, and numeric mismatch together with LLM-generated factive, structural, lexical, world-knowledge, temporal, aspectual, causal, spatial, ideological, modal, quantitative, and probabilistic mismatch types to support type-weighted and severity-aware CI constructions (Pielka et al., 2023). Enterprise contradiction detection explicitly couples automated contradiction mining with human-in-the-loop validation. Peer-review contradiction analysis uses expert annotations, evidence spans, rationale generation, and adjudication. These developments suggest that contradiction indices are increasingly treated as audit instruments rather than raw classifier scores (Mantravadi et al., 3 Oct 2025, Kumar et al., 11 May 2026).

A final unresolved issue is portability. Some papers provide structural characterization without algorithmic complexity bounds, as in the contradiction-graph test for VC threshold detection. Others provide benchmark-specific indices whose thresholds, weights, or penalties are intentionally application-dependent. The result is a family of well-defined but domain-bound contradiction indices: exact and invariant in some settings, calibrated and evidence-sensitive in others, and frequently derived rather than primitive. The common denominator is not a shared formula but a shared purpose—turning contradiction from a qualitative judgment into a reproducible scalar or ordinal object that can support detection, comparison, and decision-making.