Hallucination Guardrail

A decoding strategy that constrains a language model to synthesize text only from a specific retrieved passage. Unlike post-hoc verification, grounded generation prevents hallucination at inference time by forcing the model to condition its output on a provided evidence span. If the model cannot find support in the source, it must abstain rather than confabulate.

• Attribution mapping: Every generated claim is linked to a source span
• Abstention mechanism: Model outputs a refusal token when evidence is insufficient
• Contrast with RAG: Standard RAG retrieves context but does not enforce fidelity to it

A two-stage architecture where a cited authority is first retrieved from a ground-truth database and then programmatically compared against the model's generated summary. The system confirms factual consistency using natural language inference models that classify whether the generated text is entailed by, contradicts, or is neutral toward the source holding.

• Entailment scoring: Measures logical necessity between source and generated text
• Ground-truth database: A sealed, immutable corpus of validated legal authorities
• Pipeline order: Retrieval occurs before the user sees the output, not after

An NLP task that identifies logical inconsistencies between a generated legal proposition and the holding of the authority it purports to cite. Modern systems use transformer-based natural language inference models fine-tuned on legal contradiction datasets to flag when a model asserts the opposite of what a case actually held.

• Three-way classification: Entailment, contradiction, neutral
• Legal NLI datasets: Specialized corpora like CaseHOLD for training
• Threshold tuning: Adjusting sensitivity to balance false positives and missed fabrications

The process of using a citator service to verify the current validity and precedential weight of a legal authority. Named after Shepard's Citations, this process traces a case's subsequent judicial and legislative treatment history to determine if it remains good law. Automated Shepardizing is a critical input to hallucination guardrails, ensuring that even correctly cited authorities are still binding.

• Treatment flags: Warning symbols for overruled, questioned, or superseded cases
• History chain: Complete appellate lineage through remands and vacaturs
• Negative treatment detection: Identifying cases that have been criticized or limited

The computational process of converting diverse legal citation formats into a single canonical form. Legal citations vary widely—The Bluebook, ALWD, state-specific formats, and vendor-specific codes—and normalization enables reliable cross-database matching. Without normalization, a hallucination guardrail cannot verify whether a generated citation matches a real authority.

• Regex parsing: Pattern-based extraction of volume, reporter, and page
• Fuzzy matching: Approximate string comparison for typographical errors
• Canonical ID assignment: Mapping all variants to a unique authority identifier

An automated jurisdictional filter that determines whether a cited case originates from a higher court within the same appellate path. A hallucination guardrail must not only verify that a citation exists but also that it carries mandatory precedential weight for the legal issue at hand. Citing a persuasive authority as binding is a subtle form of legal error.

• Appellate path mapping: Tracing the hierarchy from district to circuit to Supreme Court
• Jurisdictional relevance scoring: Weighting authority by geographic and subject-matter nexus
• Stare decisis rules engine: Encoding the binding vs. persuasive distinction per jurisdiction