ClauseMark

Problem

UN digital-trade benchmarking (RDTII) requires mapping real statutes to policy indicators with defensible evidence. Legal RAG systems routinely hallucinate citations; isolated clause retrieval misses exceptions and definitions in other sections; a confident "0" without measured recall is epistemically unsafe for regulators.

Approach

Regulatory Intelligence Engine (RIE): parent-document hybrid retrieval over a legal structure graph; constrained decoding for indicator IDs; four verification gates (deterministic span checks + NLI + second LLM + self-consistency). Layer 1 is verifiable extraction; Layer 2 is a recommended score band that always requires human confirmation. Pillars and indicators live in YAML-not Python-so all 12 RDTII pillars extend by configuration.

At a glance

Tech decisions

• ID-replacement citations

  The model never authors citation strings-only span IDs resolved by deterministic code, blocking ghost references.

• Two-layer output

  Extraction is automatable and gate-checked; scoring is legal judgment and stays a reviewer-confirmed recommendation.

• Parent-document retrieval

  Child-chunk search with parent+neighbourhood context recovers cross-section exceptions standard chunking severs.

• Pillar-as-data

  Engine stays generic; legal expertise ships in validated YAML + gold sets-add a pillar without refactoring core code.

• LangGraph + Postgres checkpointer

  Durable state and interrupt-based human review for flagged claims and absence cases.