consensus-pipeline

More about this plugin

The pipeline in one sentence: Search papers → 3-layer QC filter → 11 departments debate each claim → cross-department validation → generate confidence-annotated report with NLI citation verification.

🔍 Multi-Source Paper Search & 3-Layer QC — Queries OpenAlex + Semantic Scholar + arXiv with automatic deduplication, followed by a hard filter → LLM classify → tag layer sieve ( 65% off-topic exclusion rate). - 🏛️ 11-Department Adversarial Debate — 2–4 specialized debaters per department with cross-department consensus handoff (eliminating isolated debate silos) and dynamic convergence ($CV$ + Kendall's $W$). - 📊 Per-Claim Confidence & Faithful Paraphrase — Every conclusion is tagged with supporting paper counts; strict transcription rules prevent fabricated data levels or mechanism directions. - ✅ Two-Stage NLI Citation Verification — Every claim is checked against retrieved abstracts and full texts, explicitly distinguishing citation mismatch (⚠️ wrong paper) from overstatement (✂️ wording too strong) while excluding unverifiable claims (📖 needs-fulltext / 📭 title-only) from polluting confidence scores. - 📄 Full-Text Breakpoint & Multi-Format Export — Auto-fetches OA full texts (Unpaywall/S2/OpenAlex), pauses at Phase 5.5 for interactive paywalled PDF supplementation, and exports bilingual reports (Markdown / DOCX / PDF) plus runnable Python code.

📋 Click to expand…