Deontic RAG

Every generated obligation is anchored to a specific statutory or regulatory provision retrieved from a vectorized legal corpus. The system outputs the normative statement alongside its provenance metadata—including jurisdiction, statute title, section number, and effective date—eliminating hallucinated duties.

• Source traceability: Each obligation links to the exact retrieved passage
• Jurisdictional scoping: Filters retrieval to the relevant sovereign authority
• Temporal grounding: Incorporates the effective date of the cited provision

Complex normative questions are decomposed into structured sub-queries targeting specific deontic modalities. A query like 'What must a data controller do after a breach?' is broken into:

• Obligation retrieval: 'data controller breach notification duty'
• Permission retrieval: 'data controller breach response authorized actions'
• Prohibition retrieval: 'data controller breach prohibited disclosures'

Each sub-query retrieves from a corpus pre-annotated with deontic annotation schemas, ensuring modality-specific precision.

When retrieved passages prescribe incompatible obligations, the system flags the conflict and applies resolution heuristics derived from normative hierarchy principles:

• Lex superior: Higher authority prevails (constitution over statute)
• Lex specialis: Specific provision overrides general
• Lex posterior: Later enactment supersedes earlier

The RAG pipeline surfaces the conflict explicitly rather than silently selecting one norm, enabling transparent legal analysis.

Standard RAG systems fail when primary obligations are violated. Deontic RAG incorporates contrary-to-duty (CTD) retrieval—when a breach condition is detected in the context, the system retrieves the secondary norms that govern non-ideal compliance scenarios.

• Detects violation triggers in the input fact pattern
• Retrieves remedial obligations and penalty provisions
• Models the full normative lifecycle: primary duty → breach → secondary duty

The retrieval pipeline maps legal relations into Hohfeldian jural correlatives to disambiguate normative positions:

• Right ↔ Duty: If Party A has a right, Party B has a correlative duty
• Privilege ↔ No-right: If Party A has a privilege, Party B has no claim against it
• Power ↔ Liability: If Party A has a power, Party B is liable to its exercise

This structured extraction prevents the conflation of distinct normative relations in generated outputs.

Generated outputs pass through a deontic guardrail—a runtime validation layer that checks the logical consistency of the normative conclusions before presentation:

• Verifies that no generated obligation contradicts a retrieved prohibition
• Ensures the ought-implies-can principle is not violated
• Flags internally inconsistent normative chains for human review

This layer acts as a safety net, catching errors that survive the retrieval and generation stages.

Retrieval-augmented generation systems specifically grounded in legal corpora—statutes, regulations, and case law. Unlike generic RAG, legal RAG requires:

• Citation-aware chunking that preserves section numbering and hierarchical structure
• Jurisdictional filtering to prevent cross-contamination of authority
• Temporal versioning to retrieve the law as it existed at the relevant date Deontic RAG extends legal RAG by adding a normative reasoning layer that interprets retrieved text through obligation and permission operators.

The algorithmic process of detecting and reconciling contradictory legal rules retrieved from different sources. When a RAG system pulls both a general statute and a specific exception, conflict resolution applies:

• Lex superior: Higher authority prevails
• Lex specialis: Specific rule overrides general
• Lex posterior: Later enactment controls Deontic RAG must implement these resolution strategies at inference time to avoid generating contradictory obligations.

Automated validation engines that check whether a generated legal reference corresponds to a real, correctly cited authority. In deontic RAG, every obligation must be traceable to its source. These systems:

• Validate citation format against jurisdiction-specific style guides
• Verify the continued good law status of cited authority
• Detect hallucinated citations before they reach the user Citation verification closes the loop between retrieval and generation, ensuring normative claims are auditable.

A runtime constraint mechanism that filters generative AI output to ensure it does not:

• Prescribe illegal actions
• Violate encoded normative policies
• Generate internally contradictory obligations In a deontic RAG pipeline, guardrails act as a final validation layer, comparing generated obligations against the retrieved normative framework and blocking outputs that fail consistency checks.

A quantitative evaluation score measuring how accurately generated legal text reflects the deontic content of its source material. Key dimensions include:

• Obligation recall: Are all duties from the source present?
• Permission precision: Are permissions correctly scoped?
• Prohibition fidelity: Are prohibitions neither weakened nor expanded? This metric is essential for benchmarking deontic RAG systems against human legal analysis.