Contradiction Finder

contradiction-finder — Cross-Paper Conflict Detector

Identifies genuine contradictions between claims across a set of papers. The goal is not to summarize what each paper says in isolation, but to surface pairwise conflicts: where Paper A makes a claim that Paper B directly disagrees with, quantitatively contradicts, or meaningfully qualifies.

This is different from what Elicit and Semantic Scholar do — those tools show you the papers. This skill shows you the conflicts between them.

What counts as a contradiction

Three types, in order of confidence:

DIRECT — Paper A explicitly asserts X; Paper B explicitly asserts not-X. Example: A says "the primary mechanism is cation migration"; B says "the primary mechanism is electroosmotic water flow."

QUANTITATIVE — Both papers measure the same quantity under comparable conditions but report substantially different values. Example: A reports "blocking force of 0.3 mN at 3V"; B reports "blocking force of 2.1 mN at 3V" for the same material class. Note: values differing by ≤20% may reflect measurement uncertainty rather than genuine conflict — flag but note this.

Before comparing quantities, normalize units and recognize synonymous quantities — otherwise you will miss real conflicts whenever two groups use different nomenclature (the most important case to catch). Convert to a common unit ("30 µN" = "0.03 mN") and map terms that name the same physical quantity ("blocking force" / "tip force"; "strain" / "tip deflection" / "bending angle" are related but NOT identical — only treat as comparable if they measure the same thing). If two papers appear to disagree but use different quantity names, state the mapping you applied so the user can check it.

SCOPE — A makes a general claim; B shows it holds only under specific conditions, or vice versa. Example: A says "IPMC actuators degrade in air"; B says "ionic liquid-based IPMCs operate stably in air for 180,000 cycles." This is often not a true contradiction but a boundary condition — note this distinction.

What is NOT a contradiction:

• Papers measuring different materials, fabrication methods, or test conditions (different systems, not conflicting claims)
• Papers at different scales (single-cell vs tissue-level, nanoparticle vs bulk)
• Papers from different eras where the later one has better data
• Disagreement about interpretation while agreeing on the data

Do not default to SCOPE. It is tempting to explain away every conflict as "maybe different conditions" — but that bias turns this tool into one that reports "no contradictions found" for a genuinely contested field, defeating its purpose. Apply this guardrail:

• If both papers address the same material system, same measured quantity, and comparable conditions, and their claims oppose → this is DIRECT or QUANTITATIVE, not SCOPE. The same-system + same-quantity match overrides any speculative "could be methodology" explanation.
• Only classify as SCOPE when there is a concrete, identifiable boundary-condition difference (named different material, explicitly different test environment, different scale) — not a hypothetical one you invented to dissolve the conflict.
• The "Could explain" field in the output is for genuine, checkable hypotheses — not for manufacturing excuses to downgrade every conflict. If you cannot name a specific difference, do not invent one; report the conflict at its real severity.

Protocol

Step 1 — Build the paper set

If DOIs provided: First confirm each DOI resolves to a real paper before using it. Fetch via CrossRef (https://api.crossref.org/works/{DOI}) — a 404 means the DOI does not exist, and you must not invent an abstract for it. Then fetch each abstract via CrossRef and Semantic Scholar (same protocol as ref-verify, Layer 3). If a DOI is invalid, report it and exclude it. If fewer than 3 valid papers remain, warn the user that contradiction detection requires at least 3 sources.

If search query provided: Search arXiv and Semantic Scholar for the topic, fetch top 6-8 papers sorted by citation count (established papers are more likely to represent real positions in a debate). Fetch abstracts.

Limit: 8 papers maximum. Pairwise comparison scales as N(N-1)/2 — 6 papers = 15 pairs, 8 papers = 28 pairs. A full 8-paper analysis can take 15-20+ minutes. Warn the user of the expected time before starting a set of 6+ papers, and offer a faster triage: first cluster claims by topic, then only compare within clusters (papers about different quantities rarely conflict). If more than 8 papers are provided, ask the user to select the most representative ones.

Step 2 — Extract atomic claims

From each abstract, extract specific, falsifiable claims. Focus on:

• Mechanism claims ("the mechanism is X")
• Quantitative results ("property Y = Z under conditions W")
• Causation claims ("factor A causes effect B")
• Universality claims ("always", "all", "never", "generally")

Do not extract:

• Background statements ("X is important for Y applications")
• Motivation ("this study aims to...")
• Vague claims that cannot be falsified

Mark each claim with its source paper and verbatim excerpt. Example:

[CLAIM] Paper 1 (DOI:...): "the dominant actuation mechanism is cation migration toward the cathode"
— verbatim: "our results demonstrate that Cs⁺ migration is the primary contributor to bending deformation"

Step 3 — Pairwise comparison

Compare each paper's claims against every other paper's claims. For each pair, ask:

1. Do any claims make assertions about the same physical quantity, mechanism, or relationship?
2. If yes: do they agree, conflict, or qualify each other?
3. If they conflict: is this a genuine contradiction or a methodological difference?

Step 4 — Classify and filter

Assign conflict severity:

• HIGH: Direct contradiction on a core claim, same material system, no obvious methodological explanation
• MEDIUM: Quantitative disagreement beyond measurement uncertainty, or SCOPE conflict where the scope distinction matters for practical use
• LOW: Apparent contradiction likely explained by methodology, scale, or era difference

Discard LOW-severity conflicts from the main report — list them only in an appendix.

Output Format

CONTRADICTION ANALYSIS
Papers analyzed: [N]
Date: [date]
────────────────────────────────────────────────────

HIGH-SEVERITY CONFLICTS
═══════════════════════

[1] DIRECT CONTRADICTION — [Short description]
────────────────────────────────────────────────────
Claim A: "[verbatim quote from abstract]"
  Source: [Author et al. (year)] | DOI: ... | arXiv: ...

Claim B: "[verbatim quote from abstract]"
  Source: [Author et al. (year)] | DOI: ... | arXiv: ...

Conflict: [1-2 sentence explanation of why these contradict]
Could explain: [Possible reasons: different cation types? different humidity? different electrode prep?]
────────────────────────────────────────────────────

[2] QUANTITATIVE DISAGREEMENT — [Short description]
...

MEDIUM-SEVERITY CONFLICTS
══════════════════════════
[list with abbreviated format]

SUMMARY
────────────────────────────────────────────────────
High:   [N] conflicts
Medium: [N] conflicts
Low:    [N] (listed in appendix, likely methodological)

Most contested claim: [the claim that appears in the most conflicts]
Suggested resolution: [What experiment or measurement would resolve the main conflict?]
────────────────────────────────────────────────────

Handling common pitfalls

Era mismatch: A 2005 paper and a 2024 paper may appear to contradict when the 2024 paper simply has better instrumentation. Flag the year difference and note this as a possible explanation.

Different material systems labeled similarly: "dielectric elastomer" can mean VHB, silicone, polyurethane, or custom chemistries. If two papers use the same umbrella term but different materials, the "contradiction" may not be one.

Selective abstract reporting: Authors choose what to include in abstracts. A claim absent from an abstract doesn't mean the paper disagrees — it means the paper doesn't address that claim. Do not infer disagreement from silence.

Systematic vs. anecdotal evidence: A single-experiment result vs. a meta-analysis covering 50 studies is not a fair contradiction. Note the evidence level difference.

Minimum viable output

If abstracts are available for fewer than 3 papers, or if no contradictions are found, still produce a useful output:

• List the claims extracted from each paper
• State explicitly: "No contradictions found between these papers at the abstract level. Either the papers address compatible questions, or the conflicts are in the body text (not the abstract)."
• Suggest: "To find deeper conflicts, provide full-text access or look for papers that specifically rebut each other."