Groundedness Detection

At its core, groundedness detection relies on Natural Language Inference (NLI) to classify the logical relationship between a source premise and a generated hypothesis. The system determines if a claim is entailed by (supported), contradicts (refuted), or is neutral to the source text. In legal applications, a claim that is neutral is treated as a hallucination, as every assertion must be explicitly grounded in the provided evidence.

Beyond binary classification, sophisticated systems produce an attribution score that quantifies the degree of support. This metric pinpoints the exact source passage that grounds a claim, creating a verifiable provenance trail. Key components include:

• Token-level grounding: Mapping specific words in the output to source spans
• Evidence highlighting: Visually indicating the supporting text for a human reviewer
• Confidence calibration: Ensuring the score reflects true empirical accuracy

In legal AI, groundedness is measured through dual metrics of citation recall and citation precision. Citation recall measures the proportion of factual claims that are correctly supported by a citation, while citation precision verifies that each provided citation genuinely supports its associated claim. A system with high recall but low precision may be fabricating authoritative-sounding references, a critical failure mode in legal contexts.

A strict form of groundedness detection, context adherence evaluates whether a response is derived exclusively from the user-provided context. This penalizes the introduction of external knowledge or parametric memory. Enforcement mechanisms include:

• Closed-book prohibition: Rejecting any claim not traceable to the input
• Schema-constrained decoding: Forcing outputs to conform to a predefined structure
• Abstention triggers: Programming the system to state 'insufficient information' rather than guess

When synthesizing across multiple legal documents, groundedness detection must also identify cross-document contradictions. A claim may be supported by one source but directly refuted by another. This requires multi-hop reasoning to surface logical inconsistencies across a corpus, ensuring the final output does not present a contradictory synthesis as coherent fact. This is essential for case law analysis where precedents may conflict.

A common architectural pattern employs a dedicated verifier model—a secondary, often smaller language model trained to act as a critic. This model checks the primary generator's output for factual errors before user delivery. The verifier can be:

• Fine-tuned on hallucination datasets to recognize common failure patterns
• Prompted with chain-of-verification to generate fact-checking questions
• Integrated as a filter in a fact verification pipeline that decomposes, retrieves, and adjudicates claims

The foundational classification task for automated fact-checking. NLI determines if a hypothesis (the generated claim) can be logically inferred from a premise (the source text).

• Entailment: The source text proves the claim.
• Contradiction: The source text disproves the claim.
• Neutral: The source text does not provide enough information.

In legal AI, a fine-tuned NLI model acts as the primary gatekeeper, blocking any generated sentence that does not achieve an entailment score above a strict threshold.

A quantitative metric that measures the provenance quality of a generated statement. It goes beyond binary fact-checking to calculate the degree to which a claim can be directly linked to a specific, granular segment of a source document.

• High attribution: 'The contract terminates on June 1st' is linked to Section 12(b), Paragraph 3.
• Low attribution: A summary statement that blends concepts from five different sections without clear mapping.

This score is critical for auditability, allowing a lawyer to click from an AI-generated brief directly to the supporting evidence.

A dual metric borrowed from information retrieval, adapted to measure the integrity of a legal AI's references.

• Citation Recall: The proportion of factual claims in a generated text that are correctly supported by a citation. Low recall means the model is making unsupported assertions.
• Citation Precision: The proportion of provided citations that genuinely support the associated claim. Low precision detects hallucinated references—citations to cases or statutes that do not exist or do not say what the model claims.

A perfect system achieves 100% on both, ensuring every statement has a real, correct authority.

A two-stage generation paradigm where a secondary, often smaller and faster, language model is trained exclusively as a critic. Its sole function is to check the primary model's output for factual errors and logical inconsistencies.

• The primary model generates a draft legal analysis.
• The verifier model decomposes each claim and checks it against the provided source documents.
• Any claim that fails verification is flagged for revision or deletion.

This creates a system of checks and balances, preventing a single model's hallucination from reaching the end-user.

A faithfulness metric that evaluates whether a model's response is strictly derived from the user-provided context. It penalizes the introduction of external knowledge or assumptions not present in the input.

• Example violation: When asked to summarize a contract that is silent on governing law, the model incorrectly states 'This agreement is governed by Delaware law' based on its pre-training data.
• A high context adherence score means the model acts as a perfect mirror of the provided documents, adding nothing and omitting nothing material.

This is the core principle behind 'grounded' generation.

The computational task of identifying mutually exclusive statements within a single document or across a multi-document corpus. This is critical for surfacing logical inconsistencies in legal reasoning.

• Intra-document: 'The Seller shall deliver goods on June 1st' vs. 'Delivery shall occur no later than May 15th.'
• Cross-document: A statute in one jurisdiction conflicts with a regulation in another.

Advanced groundedness systems use contradiction detection to prevent the AI from synthesizing a 'Frankenstein' argument that blends incompatible legal principles from different sources.