Grounded Generation

The core mechanism that constrains the decoder to only generate tokens that can be directly mapped to a source passage. Unlike standard generation, the model is prohibited from synthesizing text based on its parametric knowledge. Every generated statement must have a provenance pointer linking it to a specific span in the retrieved document. This is often implemented via constrained beam search or post-hoc alignment verification that filters unsupported claims before they reach the user.

Grounded generation operates on a strict two-stage architecture:

• Stage 1 - Retrieval: A dense retrieval model fetches the top-k relevant chunks from a vector store containing the authoritative legal corpus.
• Stage 2 - Synthesis: The language model generates text conditioned only on the retrieved chunks, with no access to its pre-trained weights for factual claims. This separation ensures the model cannot hallucinate by blending retrieved facts with memorized but potentially outdated information.

A post-generation evaluation metric that quantifies how faithfully the output represents the source material. Common implementations include:

• Natural Language Inference (NLI): Treats the source passage as a premise and the generated claim as a hypothesis, scoring for entailment.
• Token-level alignment: Uses attention weights or saliency methods to verify that generated tokens are grounded in the input context.
• Factual Consistency (FactCC): A dedicated model trained to detect contradictions between a source document and a generated summary. Scores below a defined threshold trigger regeneration or flag the output for human review.

Grounded generation directly addresses the extrinsic hallucination problem in legal AI. By design, the system cannot invent case names, fabricate holdings, or misattribute statutory language. The ablation is achieved through architectural constraints, not just prompt engineering:

• The model's vocabulary is effectively restricted to concepts present in the retrieved evidence.
• Knowledge cutoff is irrelevant because the model relies on the real-time retrieved corpus, not its training data.
• This provides a verifiable audit trail from every output sentence back to a citable legal authority.

A fine-grained grounding technique where the model is trained to output inline citation markers that point to exact character offsets in the source document. This goes beyond document-level attribution to provide passage-level and sentence-level provenance. Training data for this capability is created by automatically aligning existing legal summaries with their cited sources and labeling the minimal text span that supports each statement. The output format often uses XML-style tags or structured JSON to demarcate the evidence boundaries.

While Retrieval-Augmented Generation (RAG) provides context to a model, grounded generation imposes a stricter contract:

• RAG: The model may use the retrieved context, but can still draw on parametric knowledge.
• Grounded Generation: The model must use the retrieved context and is explicitly prevented from extrapolating. This distinction is critical in legal domains where citation integrity is non-negotiable. Grounded generation is essentially RAG with a zero-tolerance hallucination policy enforced at the architectural level.

A system architecture that programmatically confirms a model's generated summary is factually consistent with the source text. The process first retrieves a cited authority from a ground-truth database, then uses a Natural Language Inference (NLI) model to detect contradictions between the generated proposition and the original holding. This acts as a critical safety net, preventing the dissemination of plausible-sounding but fabricated legal interpretations.

A verification layer in legal AI systems that intercepts generated text to detect and suppress fabricated case names, citations, or holdings before they reach the user. Key techniques include:

• Schema-constrained decoding: Forcing the model to only generate tokens that match a valid citation format.
• Entity resolution: Cross-referencing generated entities against a knowledge base of real cases.
• Logit manipulation: Penalizing the generation of unknown or statistically unlikely citation strings.

An NLP task that identifies logical inconsistencies between a generated legal proposition and the holding of the authority it purports to cite. This goes beyond simple string matching by employing Natural Language Inference (NLI) models fine-tuned on legal text. The system classifies the relationship between the source text and the generated summary as entailment, contradiction, or neutral, flagging any output that misrepresents the court's actual reasoning.

The surrounding textual passage analyzed alongside a citation to determine the author's intent. A citation can be followed, distinguished, criticized, or overruled. By analyzing the context window, a grounded generation system can understand whether the cited authority is being used as positive support or as a counterpoint, preventing the model from incorrectly treating a criticized case as binding precedent.

A composite algorithmic ranking of a legal citation's value based on a weighted combination of factors:

• Court Level: Supreme Court vs. district court.
• Case Age: Recency of the decision.
• Depth of Treatment: How substantively the case was analyzed by subsequent courts.
• Negative History: Whether the case has been criticized or overruled. This score guides the retrieval system to prioritize the most authoritative sources for grounding generation.

The process of using a citator service like Shepard's Citations to verify the current validity and precedential weight of a legal authority. It traces the subsequent judicial and legislative treatment history of a case to determine if it is still good law. A grounded generation system integrates this signal to ensure it never synthesizes a legal argument based on an authority that has been implicitly overruled or superseded.