CLAIM GRAPH: Autonomous Evidence & Contradiction Intelligence Engine

Описание и применение

CLAIM GRAPH автоматически преобразует неструктурированные документы в граф знаний, отслеживает временные изменения и выявляет недействительные выводы, снижая риск ошибок в инженерных проектах.

Задача

Teams in engineering, research, and compliance constantly revise documents, causing downstream conclusions to rely on outdated, refuted, or superseded premises. Traditional search and LLM chat interfaces cannot maintain strict logical dependency graphs across document versions, leading to costly design flaws and contradictory assumptions.

Результат

CLAIM GRAPH provides an autonomous, real‑time evidence and contradiction intelligence engine that transforms unstructured documents into a testable knowledge graph, automatically tracks temporal changes, and surfaces invalidated conclusions, thereby reducing design errors and improving compliance.

Рабочий процесс

• 1. Upload or link unstructured documents (specs, papers, policy briefs).
• 2. Engine parses text into atomic claims classified as FACT, SPECIFICATION, CONCLUSION, or ASSUMPTION.
• 3. Claims are mapped into a NetworkX directed property graph with edges (Premise) -[:DERIVES]-> (Conclusion).
• 4. Temporal analysis compares document revisions; older claims are marked SUPERSEDED or CONTRADICTED based on date and conflict.
• 5. Orphaned conclusions are automatically flagged when their premises are invalidated.
• 6. Users view the interactive physics‑simulated graph, status matrices, and watchlists to review and resolve issues.

Доказательства

CLAIM GRAPH is an autonomous evidence & contradiction intelligence engine that deconstructs unstructured architectural specifications, research papers, and policy briefs into testable atomic claims.

Источник

It dynamically maps dependencies, reconciles cross-document temporal revisions, and pinpoints invalidated conclusions whose supporting evidence has fallen.

Источник

Atomic Claim Extraction: Parses unstructured document chunks into discrete, testable propositions classified as FACT, SPECIFICATION, CONCLUSION, or ASSUMPTION.

Источник

Temporal Contradiction Tracking: Detects direct conflicts between specifications across document revisions and chronologically marks older claims as SUPERSEDED and same-date conflicts as CONTRADICTED.

Источник

Orphan & Invalidation Analysis: Constructs directed dependency edges (Premise) -[:DERIVES]-> (Conclusion). If a premise is superseded or contradicted, the dependent conclusion is automatically invalidated as ORPHANED.

Источник

Built with Python, FastAPI, NetworkX, Pydantic, and Ollama (gemma3:1b), orchestrated with Antigravity AI coding assistant and verified via automated E2E test runs.

Источник