Aggregator Contradiction

• Aggregator contradiction is the phenomenon where locally consistent inputs result in globally inconsistent or inefficient outcomes across domains such as binary voting, argumentation, and RAG.
• It is characterized by failures of aggregation rules, as safe designs (e.g., 2-CNF constraints) maintain collective rationality while higher-arity dependencies often lead to paradoxes.
• Design responses include tailored aggregation methods, contradiction detection algorithms, and human-in-the-loop validation to ensure that collective outputs align with underlying semantic or logical constraints.

Searching arXiv for the cited papers to ground the response in current records. Aggregator contradiction denotes a family of aggregation failures in which acceptable local inputs do not compose into an acceptable global output. In binary aggregation with integrity constraints, it is the case where every individual ballot satisfies an integrity constraint $\phi$ but the aggregated ballot does not (Grandi, 2014). In Retrieval-Augmented Generation (RAG), it is a context-context conflict inside the retrieved set of documents (Gokul et al., 31 Mar 2025). In multi-agent argumentation, it appears as failure of collective rationality or as non-preservation of semantic properties under Arrovian aggregation [(Awad et al., 2014); (Chen et al., 2017)]. In forecast aggregation, the contradiction is not logical inconsistency but the result that any aggregator that always remains strictly between the smallest and largest forecasts is never efficient in practice (Satopää, 2017). In agent-based dynamics, it also names the tension between a naïve macro-aggregation and the behavior of a more faithful heterogeneous model (Banisch, 2015).

1. Cross-domain meaning and scope

The literature uses the expression in several precise senses, each tied to a specific aggregation problem.

Domain	Aggregate object	Contradiction
Binary aggregation	Collective ballot	$F(B)\notin \mathrm{Mod}(\phi)$
Judgment aggregation in argumentation	Social labelling	$F(L)\notin \mathrm{Comp}(AF)$
AF aggregation	Collective attack relation	Failure to preserve an AF-property
RAG and enterprise RAG	Retrieved document set	Mutually incompatible statements
Social-media rumor analysis	Tweet pairs	Contradictory claims about the same claim-target
Forecasting and macro-dynamics	Aggregate prediction or macro-state	Inefficiency or qualitative mismatch

In binary aggregation, the contradiction is formalized as a paradox. In argumentation, it is a violation of collective rationality or of preservation. In text-centric systems such as rumor analysis and RAG, it becomes a detection problem over aggregated evidence. In forecasting and dynamical systems, it concerns the incompatibility between a central-tendency aggregate and the information or dynamics that the aggregate is supposed to represent. This suggests that the term spans both normative impossibility results and operational consistency failures.

2. Binary aggregation with integrity constraints

The most explicit general definition is given in the framework of binary aggregation with integrity constraints. Let $\Pi=\{p_1,\dots,p_m\}$ be an agenda of yes/no issues, let an individual ballot be a function $b_i:\Pi\to\{0,1\}$, and let an integrity constraint $\phi$ be an arbitrary propositional formula over $\Pi$ selecting the rational ballots,

$$\mathrm{Mod}(\phi)=\{B\in\{0,1\}^m:B\models \phi\}.$$

A profile is a list of individual ballots $B=(B_1,\dots,B_n)$, and an aggregator $F$ maps profiles into collective ballots. Under simple majority,

$F(B)\notin \mathrm{Mod}(\phi)$0

A triple $F(B)\notin \mathrm{Mod}(\phi)$1 is a paradox exactly when every individual ballot in $F(B)\notin \mathrm{Mod}(\phi)$2 satisfies $F(B)\notin \mathrm{Mod}(\phi)$3 but the collective outcome fails to satisfy $F(B)\notin \mathrm{Mod}(\phi)$4, that is,

$F(B)\notin \mathrm{Mod}(\phi)$5

This formulation subsumes the Condorcet paradox, the discursive dilemma, and the Ostrogorski paradox (Grandi, 2014).

The Condorcet paradox is obtained by taking $F(B)\notin \mathrm{Mod}(\phi)$6 and letting $F(B)\notin \mathrm{Mod}(\phi)$7 encode transitivity, for instance

$F(B)\notin \mathrm{Mod}(\phi)$8

The discursive dilemma uses $F(B)\notin \mathrm{Mod}(\phi)$9 with a consistency constraint, and the Ostrogorski paradox uses $F(L)\notin \mathrm{Comp}(AF)$0 with

$F(L)\notin \mathrm{Comp}(AF)$1

In each case, issue-wise aggregation preserves local judgments while violating a higher-arity logical dependency among issues.

The main characterization isolates the syntactic source of contradiction. Whenever $F(L)\notin \mathrm{Comp}(AF)$2 has a prime-implicate of size at least $F(L)\notin \mathrm{Comp}(AF)$3, the majority rule can produce a paradox. Conversely, simple majority with odd $F(L)\notin \mathrm{Comp}(AF)$4 is collectively rational with respect to $F(L)\notin \mathrm{Comp}(AF)$5 if and only if $F(L)\notin \mathrm{Comp}(AF)$6 is equivalent to a conjunction of clauses of size at most $F(L)\notin \mathrm{Comp}(AF)$7; equivalently, the safe constraints are exactly the $F(L)\notin \mathrm{Comp}(AF)$8-CNF formulas. The boundary is therefore sharp: $F(L)\notin \mathrm{Comp}(AF)$9-CNF is safe, while any $\Pi=\{p_1,\dots,p_m\}$0-clause with $\Pi=\{p_1,\dots,p_m\}$1 is unsafe under majority. The paper also lists implication chains $\Pi=\{p_1,\dots,p_m\}$2 (clause $\Pi=\{p_1,\dots,p_m\}$3), at-most-one constraints $\Pi=\{p_1,\dots,p_m\}$4, and simple “if either $\Pi=\{p_1,\dots,p_m\}$5 or $\Pi=\{p_1,\dots,p_m\}$6 then $\Pi=\{p_1,\dots,p_m\}$7” constraints as examples of safe formulas. Typical requirements such as transitivity are not safe.

3. Judgment aggregation and abstract argumentation

In multi-agent argumentation, aggregator contradiction is formulated through collective rationality. An abstract argumentation framework is a pair

$\Pi=\{p_1,\dots,p_m\}$8

where $\Pi=\{p_1,\dots,p_m\}$9 is a finite set of arguments and $b_i:\Pi\to\{0,1\}$0 is the defeat-relation. A labelling is a total function

$b_i:\Pi\to\{0,1\}$1

and a complete labelling must satisfy the legality conditions for accepted, rejected, and undecided arguments. Given $b_i:\Pi\to\{0,1\}$2 agents, each reports a complete labelling $b_i:\Pi\to\{0,1\}$3, and an aggregation operator maps a labelling profile to a collective labelling. Argument-wise Plurality Rule chooses, for each argument, the label that appears strictly more often than any other, when no ties occur. The contradiction arises when an operator satisfies desiderata such as Anonymity and Independence but fails Collective-Rationality, so that $b_i:\Pi\to\{0,1\}$4. The paper proves several impossibility theorems showing that combinations of Universal-Domain, Systematicity, Anonymity, Collective Rationality, and Unanimity cannot coexist in general, though positive escape results exist under domain restriction to a unique semantics or graph-theoretic restrictions such as disconnected issues or limited defeaters (Awad et al., 2014).

A related line studies aggregation of the argumentation frameworks themselves rather than aggregation of labellings. Here each agent supplies an attack relation on a fixed set of arguments, and an aggregation operator returns a collective attack relation. The central question is preservation: if every individual framework satisfies an AF-property $b_i:\Pi\to\{0,1\}$5, must the collective framework satisfy $b_i:\Pi\to\{0,1\}$6 as well? The paper defines Arrovian-style axioms including Independence, Neutrality, Unanimity, Groundedness, Monotonicity, and Anonymity, and compares quota rules, oligarchies, and dictatorships. Positive results include: every grounded and unanimous operator preserves conflict-freeness; the nomination rule preserves admissibility of every set $b_i:\Pi\to\{0,1\}$7 and preserves “$b_i:\Pi\to\{0,1\}$8 is a stable extension”; and the unanimity rule preserves nonemptiness of the grounded extension. Negative results are substantially stronger: for argument acceptability under grounded, stable, preferred, or complete semantics, any operator satisfying Unanimity, Groundedness, Neutrality, and Independence and preserving acceptability must be a dictatorship when $b_i:\Pi\to\{0,1\}$9; preserving the exact grounded extension is dictatorial when $\phi$0; preserving nonempty grounded extension or acyclicity forces veto power; and preserving coherence is again dictatorial (Chen et al., 2017).

These results place argumentation-theoretic aggregator contradiction within the standard impossibility tradition. The contradiction is not merely a pathological example but a structural incompatibility between fairness, independence, and preservation of semantic legality.

4. Contradiction in aggregated textual evidence

In text aggregation, the problem shifts from designing a rule that always preserves rationality to detecting contradictions within aggregated evidence. For social media rumors, contradiction is modeled in two settings. In the broader setting of independently posted tweets, an aggregator contradiction arises when two tweets about the same claim-target cannot simultaneously be true and neither is in reply to the other. In threaded conversations, disagreement is defined relative to a source tweet and a reply with opposite polarity. Both settings are cast as a three-way Recognizing Textual Entailment task with labels Entailment, Contradiction, and Unknown. The features are intentionally simple: cosine and $\phi$1 overlap on content-word stems and on POS tags, plus two Smith–Waterman local-alignment features, $\phi$2 and $\phi$3. On the balanced iPosts dataset, the reported mean results over four folds are Accuracy $\phi$4 and weighted $\phi$5 $\phi$6 for Random Forest, and Accuracy $\phi$7 and weighted $\phi$8 $\phi$9 for Nearest Centroid; on Threads, Random Forest reaches Accuracy $\Pi$0 and weighted $\Pi$1 $\Pi$2 (Lendvai et al., 2016).

For RAG systems, an aggregator contradiction is defined over a retrieved set $\Pi$3 and exists exactly when there are statements drawn from $\Pi$4 that cannot simultaneously hold true. The study simulates three contradiction types: self-contradiction within one document, pair contradiction between two documents, and conditional contradiction in which a third document makes two otherwise compatible documents mutually exclusive. Its generator produces a balanced dataset of $\Pi$5 samples, with $\Pi$6 no-contradiction, $\Pi$7 self, $\Pi$8 pair, and $\Pi$9 conditional. Conflict-detection is evaluated with Claude-3 Sonnet, Claude-3 Haiku, Llama-3.3 70B, and Llama-3.1 8B under Basic and Chain-of-Thought prompting. On the full dataset, the best reported configuration is Claude-3 Sonnet + CoT with Accuracy $$\mathrm{Mod}(\phi)=\{B\in\{0,1\}^m:B\models \phi\}.$$0, Precision $$\mathrm{Mod}(\phi)=\{B\in\{0,1\}^m:B\models \phi\}.$$1, Recall $$\mathrm{Mod}(\phi)=\{B\in\{0,1\}^m:B\models \phi\}.$$2, and $$\mathrm{Mod}(\phi)=\{B\in\{0,1\}^m:B\models \phi\}.$$3 $$\mathrm{Mod}(\phi)=\{B\in\{0,1\}^m:B\models \phi\}.$$4. Pair contradictions are easiest, conditional contradictions next, and self-contradiction is hardest; all models show very high precision $$\mathrm{Mod}(\phi)=\{B\in\{0,1\}^m:B\models \phi\}.$$5 but low recall $$\mathrm{Mod}(\phi)=\{B\in\{0,1\}^m:B\models \phi\}.$$6 (Gokul et al., 31 Mar 2025).

Enterprise RAG extends the problem to long-form business documents. ContraGen models contradictions along two axes: structural scope, with intra-document and cross-document contradictions, and semantic category, with Temporal, Numerical, Authority, Process, Policy-Reversal, and Specificity contradictions. Content generation uses a three-agent pipeline, Pyro4 RPC calls, metadata-driven document creation, and perplexity-based fluency control with thresholds

$$\mathrm{Mod}(\phi)=\{B\in\{0,1\}^m:B\models \phi\}.$$7

Contradiction mining combines Sentence-BERT filtering, BART-MNLI classification, GPT-4o judgment, and a confidence-weighted hybrid score with threshold $$\mathrm{Mod}(\phi)=\{B\in\{0,1\}^m:B\models \phi\}.$$8. Human-in-the-loop validation reports, for self-contradictions $$\mathrm{Mod}(\phi)=\{B\in\{0,1\}^m:B\models \phi\}.$$9, $B=(B_1,\dots,B_n)$0 and $B=(B_1,\dots,B_n)$1, and for pairwise contradictions $B=(B_1,\dots,B_n)$2, $B=(B_1,\dots,B_n)$3 and $B=(B_1,\dots,B_n)$4. The reported Hybrid $B=(B_1,\dots,B_n)$5 is $B=(B_1,\dots,B_n)$6 for self and $B=(B_1,\dots,B_n)$7 for pairwise contradictions (Mantravadi et al., 3 Oct 2025).

Across these text-centric settings, contradiction is an operational property of aggregated evidence sets rather than a failure of majority rationality. The common technical problem is that aggregation enlarges the evidence base faster than simple local consistency checks can validate it.

5. Information aggregation, inefficiency, and macro-level mismatch

In forecasting, the contradiction concerns efficiency rather than logical inconsistency. Let $B=(B_1,\dots,B_n)$8 be the future outcome, let $B=(B_1,\dots,B_n)$9 be the $F$0th forecaster’s point forecast, and assume calibration,

$F$1

An aggregator is efficient if

$F$2

where $F$3. A strict mean is any central-tendency aggregator $F$4 such that

$F$5

The main theorem states that, under finite $F$6 and positive probability of disagreement, every strict mean fails to be efficient. The practical implication given in the paper is that means, medians, trimmed means, and related central-tendency rules systematically shrink the aggregate toward the center and fail to capture all available information when forecasters bring genuinely different information sets (Satopää, 2017).

The Contrarian Voter Model gives a different version of aggregator contradiction. At the microscopic level, $F$7 agents carry binary opinions and update by imitation or contrarian behavior with rate $F$8. On the complete graph, aggregation by the global count of one opinion, $F$9, is lumpable and Markovian, with a birth–death chain and stationary distribution $F(B)\notin \mathrm{Mod}(\phi)$00; at $F(B)\notin \mathrm{Mod}(\phi)$01 the chain is doubly stochastic and $F(B)\notin \mathrm{Mod}(\phi)$02. On a two-community graph with weak coupling $F(B)\notin \mathrm{Mod}(\phi)$03, the meso-level chain on $F(B)\notin \mathrm{Mod}(\phi)$04 remains exactly Markovian, but the further projection to $F(B)\notin \mathrm{Mod}(\phi)$05 is no longer Markovian because the transition probability depends on which $F(B)\notin \mathrm{Mod}(\phi)$06 state generated the same macro-count. The paper quantifies this memory with information-theoretic measures $F(B)\notin \mathrm{Mod}(\phi)$07 and $F(B)\notin \mathrm{Mod}(\phi)$08, and reports that $F(B)\notin \mathrm{Mod}(\phi)$09 whenever $F(B)\notin \mathrm{Mod}(\phi)$10 except the trivial $F(B)\notin \mathrm{Mod}(\phi)$11 case. In this setting, a homogeneous-mixing macro-aggregation predicts one stationary pattern, while the heterogeneous model can produce long-lived polarized states and elevated mass near $F(B)\notin \mathrm{Mod}(\phi)$12 or $F(B)\notin \mathrm{Mod}(\phi)$13 (Banisch, 2015).

These examples broaden the concept. The contradiction may lie between the aggregate and a logical constraint, between the aggregate and semantic legality, between the aggregate and the full information set, or between the aggregate and the true macro-dynamics induced by heterogeneous microstructure.

6. Localization, safe aggregation, and design responses

Recent work on preference aggregation emphasizes localization rather than only detection of failure. A graph-theoretic framework based on discrete sheaves models voters as vertices of a graph $F(B)\notin \mathrm{Mod}(\phi)$14 and assigns to each vertex $F(B)\notin \mathrm{Mod}(\phi)$15 the stalk $F(B)\notin \mathrm{Mod}(\phi)$16 of total orders on the alternatives visible to that voter, and to each edge $F(B)\notin \mathrm{Mod}(\phi)$17 the stalk $F(B)\notin \mathrm{Mod}(\phi)$18. For a profile $F(B)\notin \mathrm{Mod}(\phi)$19, the edge-Obstruction Locus is

$F(B)\notin \mathrm{Mod}(\phi)$20

with per-edge indicator $F(B)\notin \mathrm{Mod}(\phi)$21 and total incompatibility index $F(B)\notin \mathrm{Mod}(\phi)$22. The same framework defines a pushforward under graph quotients via a constraint DAG: if a merge creates a directed cycle among pairwise ordering constraints, the pushforward stalk is empty. The worked Condorcet triangle example on $F(B)\notin \mathrm{Mod}(\phi)$23 shows that three distributed edge conflicts can condense into a single local impossibility, an empty stalk, after merging two voters (Sargsyan, 2 Dec 2025).

The design implications across the literature are consistent. If issue-by-issue majority must always satisfy an integrity constraint, the constraint must be $F(B)\notin \mathrm{Mod}(\phi)$24-CNF; otherwise one must weaken the constraint to a $F(B)\notin \mathrm{Mod}(\phi)$25-CNF approximation or switch to a more sophisticated aggregator such as a distance-based judgment aggregator or some weighted rule (Grandi, 2014). In argumentation, one must either weaken collective rationality, restrict the domain of votes, or restrict the structure of the framework; otherwise impossibility theorems force dictatorship, veto power, or failure of preservation (Chen et al., 2017). In RAG, holistic single-shot validation over the entire retrieved set avoids pairwise $F(B)\notin \mathrm{Mod}(\phi)$26 LLM cost while still detecting self, pair, and conditional conflicts, and prompting strategy should be adapted to model architecture (Gokul et al., 31 Mar 2025). In enterprise settings, contradiction detection is embedded in governance through confidence-weighted scoring, human adjudication, and logging of timestamps, evidence pairs, model confidences, and adjudication outcomes (Mantravadi et al., 3 Oct 2025).

A plausible implication is that aggregator contradiction marks a boundary condition for aggregation design. Whenever aggregation is performed pointwise, independently, or under a coarse state description, higher-arity constraints, semantic dependencies, heterogeneous information, or latent structure can reappear as contradiction at the aggregate level. The supplied literature identifies several remedies, but it does not eliminate the underlying pattern: aggregation is safe only when the representation, rule, and preservation target are jointly aligned.