Fact Verification Pipeline

The initial stage where a complex, multi-faceted claim is parsed into its atomic, independently verifiable sub-claims. This process uses syntactic dependency parsing and semantic role labeling to isolate individual factual assertions. For example, the statement 'The court ruled on July 5, 2023, that the patent was invalid due to prior art' is decomposed into: (1) The court issued a ruling, (2) The ruling date was July 5, 2023, (3) The subject was a patent, (4) The ruling was invalidity, (5) The grounds were prior art. Each atomic claim is then routed to the evidence retrieval stage.

A dense retrieval stage that queries a trusted, immutable corpus—such as a vector database of case law or a legal knowledge graph—for documents relevant to each atomic claim. This typically employs a hybrid search strategy combining semantic similarity (bi-encoder embeddings) with lexical matching (BM25) to ensure high recall. The system retrieves the top-k candidate passages that could potentially support or refute the claim, prioritizing authoritative primary sources over secondary commentary.

The core reasoning stage where a fine-tuned NLI model classifies the logical relationship between the retrieved evidence (the premise) and the atomic claim (the hypothesis). The model outputs one of three labels:

• Entailment: The evidence logically supports the claim.
• Contradiction: The evidence directly refutes the claim.
• Neutral: The evidence is insufficient to verify or refute the claim. In legal contexts, a neutral verdict often triggers an escalation to human review, as it may indicate an unsupported assertion or a hallucination.

The final stage that synthesizes the NLI verdicts from all atomic claims into a single, holistic assessment of the original statement. A strict aggregation policy requires all sub-claims to be entailed for a 'Supported' verdict; a single contradiction or neutral finding can render the entire statement 'Unsupported' or 'Partially Supported'. This stage also generates a structured audit trail, linking each atomic claim to its specific source passage and entailment score, providing full source attribution and enabling downstream citation verification.

A critical control mechanism that prevents the pipeline from generating a low-confidence verdict. Using uncertainty quantification techniques, the system calculates a confidence score for each NLI classification. If the model's confidence falls below a predefined threshold—or if the retrieval stage fails to find any relevant evidence—the pipeline abstains from rendering a verdict. Instead, it flags the claim for mandatory human review, ensuring that the system fails safely rather than silently propagating a hallucinated or unverified assertion.

The Fact Verification Pipeline operates as a post-hoc guardrail within a broader Retrieval-Augmented Generation (RAG) system. After a language model generates a response, the pipeline verifies every factual claim against the same retrieval corpus used for generation. This creates a closed-loop system where:

• The generator produces a draft with citations.
• The verifier pipeline checks each citation's entailment.
• Any unsupported claim is flagged for self-refine or human correction. This architecture directly addresses the high citation precision and citation recall requirements of legal AI.