Skip to content
AL | Apatheia Labs
K

Contradiction

Cross-Document Inconsistency Detection

Systematically detects inconsistencies across your document corpus using the CASCADE 8-type framework. Surfaces conflicts that undermine institutional conclusions.

CASCADE DetectionSource ComparisonConflict MappingConfidence ScoringSeverity Rating

The Problem

Document 12 says "no concerns." Document 47 says "significant concerns." Both are authored by the same professional, six weeks apart, with no intervening events to explain the shift. Buried in a 500-page corpus, this contradiction is invisible to manual review. Multiply that across hundreds of claims and dozens of documents, and systematic inconsistency becomes undetectable without automation.

How It Works

  1. 1Scan each document for internal contradictions (SELF)
  2. 2Cross-reference claims across documents (INTER_DOC)
  3. 3Validate timeline consistency (TEMPORAL)
  4. 4Check claims against cited evidence (EVIDENTIARY)
  5. 5Track certainty shifts (MODALITY)

Inputs

  • • Document corpus
  • • Claim registry
  • • Citation network

Outputs

  • • Contradiction findings
  • • Confidence/severity scores
  • • Resolution suggestions

What You Get

CASCADE Type: UNEXPLAINED_CHANGE | Severity: HIGH | Confidence: 0.94

Statement A:
  "No safeguarding concerns were raised during the visit."
  Source: Doc 12, p.3, para 2 | Author: Social Worker A | Date: 14 March 2023

Statement B:
  "Significant safeguarding concerns persist and have been present throughout the period of involvement."
  Source: Doc 47, p.8, para 1 | Author: Social Worker A | Date: 28 April 2023

Analysis: Same author, 45 days apart. No intervening events documented. No explanation for position change found in corpus. Direction: Escalation without evidential basis.

Works With

EEntity Resolution

Ensures contradictions are correctly attributed — confirming whether conflicting statements came from the same person or different individuals.

TTemporal Parser

Provides the timeline data needed to detect TEMPORAL contradictions and to sequence when positions changed.

BBias Detection

Consumes contradiction findings to determine whether conflicts follow a systematic directional pattern.

AArgumentation

Uses contradiction data to assess claim strength — a claim contradicted by the same author’s earlier statement scores lower.

Use Cases

Multi-agency investigation analysis

When police, social services, and health agencies produce separate reports about the same events, the Contradiction Engine surfaces where their accounts diverge and classifies the type of inconsistency.

Contract dispute documentation

Tracking how representations change across pre-contractual correspondence, the contract itself, and post-execution communications to identify where parties’ positions shifted.

Regulatory filing review

Comparing statements made to different regulatory bodies about the same activities, detecting where organisations told different stories to different audiences.

Technical Approach

  • Claim extraction using semantic parsing to identify propositional content — what each sentence actually asserts, independent of hedging language
  • Semantic similarity matching via sentence embeddings to find claims about the same subject across documents, even when phrased differently
  • Logical compatibility testing to determine whether two claims can coexist (compatible, contradictory, or in tension)
  • CASCADE classification assigns each finding to one of 8 types, with severity scoring based on source authority, temporal proximity, and directional favouring