LLM document-extraction systems miss a 50-year-old computer science technique for handling interdependent legal clauses

The standard LLM document-extraction pattern—extract clauses from each document independently, then apply them—produces silently wrong answers when clauses reference shared fields that other clauses modify. Example: a management-fee cap anchored to "end of investment period" and a fee waiver anchored to the same date both shift when an amendment changes that boundary date by 18 months, but extract-then-apply applies clauses in document order without detecting the interdependency.

The correct approach is fixed-point evaluation: treat clauses as typed instructions (SET, ADJUST, CONSTRAIN, GATE actions) that read and write to a shared timeline of fund parameters, then loop until applying all clauses produces no further changes to the timeline. This pattern has been standard in computer science for at least fifty years (Datalog since the 1970s, constraint-propagation networks, build-system DAGs, reactive frameworks).

LLM-driven document-AI products in 2026 do not use this technique. They treat clauses as independent entities to extract and write to a database, even though the underlying problem for legal documents is global clause evaluation, not local passage retrieval.