yzxoi/paperly-codex/contradiction-detection

Contradiction Detection

Overview

Scans a corpus of papers for conflicting empirical claims, methodological disagreements, or opposing conclusions on the same topic. Surfaces genuine scientific contradictions that the team needs to be aware of — before they cite conflicting work or build on a shaky premise.

When to Use

• User asks "do any of our papers disagree with each other?"
• User is writing a discussion section and needs to address conflicting findings
• User wants to know if a claim they're making is contested in the literature
• Lab is onboarding to a new area and needs to understand where debates lie
• User asks "why do paper A and paper B get different results on [benchmark]?"

Key Capabilities

• Extract quantitative claims and conclusions from paper full-text or abstracts
• Compare claims across papers on the same topic, dataset, or research question
• Classify contradictions by type: empirical (different results), methodological (different best practices), interpretive (different explanations for the same data)
• Score contradiction severity: minor discrepancy vs. fundamental disagreement
• Trace contradictions to potential causes: different datasets, evaluation protocols, hyperparameters
• Generate a contradiction report structured by topic cluster

Usage Examples

Scan team corpus for contradictions

contradictions = contradiction_detection.scan(
    corpus=review.get_papers(),
    topic="learning rate schedules in transformer training",
    min_confidence=0.7
)

Check if a specific claim is contested

contradiction_detection.check_claim(
    claim="Dropout consistently improves generalization in large language models",
    corpus=review.get_papers(),
    search_external=True
)

Generate contradiction report for discussion section

contradiction_detection.report(
    topic="benchmark evaluation of protein structure prediction",
    format="discussion_section_draft",
    include_resolution_suggestions=True
)

Output Format

Returns list of contradiction pairs with: paper A, paper B, conflicting claims, contradiction type, severity score, and suggested resolution or explanation. Optionally formatted as a discussion section narrative.

Best Paired With

• claim_auditor when a contribution claim may be contested
• related_work_scout when the paper needs a balanced discussion of conflicting prior work

Notes

• Works best on papers with structured results sections and explicit quantitative claims
• Interpretive contradictions are harder to detect automatically — flag for human review
• Combine with evidence-grading to assess which side of a contradiction has stronger support