RSCEv1.0.0
PubMed Literature Synthesis & Conflict Resolution

Resolve Contradictions in
Medical Literature.

Enter a research question to automatically crawl PubMed, extract clinical claims verified by source quotes, and map conflicting findings in an interactive consensus network.

1Crawl PubMed

Searches and retrieves the latest relevant peer-reviewed studies.

2Extract & Ground

AI extracts clinical findings, verified back to exact quote anchors.

3Map Conflicts

Detects contradictions using Neural NLI and LLM-based scope validation.

4Consensus Report

Synthesizes a visual network and a citation-backed consensus summary.

Suggested Queries (Click to try)
Query Bias
Provide an assertion to bias the analysis towards finding evidence that supports or refutes it.
PubMed Search Constraints
More papers lead to richer claim graphs but longer fetch cycles.
25

Pre-loaded Sandbox Demos (Loads Instantly)

Click below to instantly explore fully processed claim contradiction graphs and reports without waiting for the live PubMed crawl or consuming API tokens.

Behind the Pipeline

RSCE operates a multi-stage validation system designed to bypass superficial semantic matches and extract actual logical contradiction.

01

Ingestion & Fuzzy Verification

Crawls PubMed abstracts or PMC full-text XML. LLM extracts claims which are checked via fuzzy quote-anchor matching against the source text to prevent hallucinations.

02

Vector Search & NLI Cross-Encoder

Claims are embedded via Sentence Transformers and indexed in FAISS. High-similarity pairs are evaluated with a DeBERTa NLI cross-encoder model to identify conflicting assertions.

03

LLM Genuineness Judge

Pairs flagged by the NLI model are evaluated by Gemini 2.5 Flash to filter out false positives caused by differing experimental scopes (e.g. human trials vs. in-vitro animal studies).

04

Graph Synthesis & Citation RAG

Entities are normalized to MeSH IDs. A NetworkX claim-evidence graph is built. A final summary report is drafted by the LLM, referencing citations that are validated back to original papers.

87.1%
SciFact Precision
Validated on scientific claim-evidence benchmark
100%
Citation Grounding
Zero hallucinated references in narrative reports
~30 min
Processing Latency
For a full 25-paper run (CPU-bound NLI scoring)