Sweep thousands of leases for compliance using Amazon Quick and the Adjudicated Query pattern

The Adjudicated Query pattern introduces a method for applying generative AI in high-stakes compliance domains, such as lease auditing, by separating the natural language interaction from the deterministic decision-making process. The pattern leverages a bounded conversational layer that translates user questions into calls against a fixed set of rules stored in a separate, non-AI rules engine. This architecture aims to solve the challenges of provable completeness and defensibility that plague traditional RAG or text-to-SQL approaches for compliance work.
The system relies on an Amazon Quick interface for conversational access and an Amazon Quick Sight dashboard for full data inspection. The core guarantee of the pattern is established through a "completeness receipt," which asserts that all records in a population have been accounted for (compliant, in-breach, ambiguous, unreadable). This receipt is computed by the rules engine before any results are presented, ensuring a mathematically verifiable completeness claim.
The system separates information flow to maintain integrity: the conversational path allows natural language interaction and result narration, while the underlying data store remains strictly governed by deterministic rules. The architecture carefully restricts the generative model's role to exploratory tasks only, preventing it from making final compliance determinations.

* A portfolio operator holds 50,000 leases across multiple states.
* State landlord-tenant statutes change based on legislative schedules.
* Compliance requires determining which leases are out of line when regulations change.
* The problem demands provable completeness and defensibility for compliance checks.
* Similarity search (RAG) cannot guarantee completeness or defendability.
* Text-to-SQL risks hallucination by silently narrowing the population.
* The Adjudicated Query pattern uses a bounded conversational layer over a deterministic rules engine.
* The rules engine contains only generic comparison operators and versioned rulebook data.
* A compliance sweep produces a completeness receipt asserting that compliant + in-breach + ambiguous + unreadable equals scanned.
* Answering requires separating the model's role (narration) from the decision process (rules engine).
* The reference architecture uses Amazon Aurora Serverless v2 for storage and AWS Lambda for hosting the rules engine.
* The pattern limits model interaction to exploratory clause-search paths via specific tools.

The core tension in this design is balancing the utility of natural language accessibility against the necessity of mathematical certainty in compliance contexts. The pattern explicitly rejects generative AI's tendency toward probabilistic reasoning when dealing with liabilities, instead constraining the LLM to a purely interpretive role. This constraint—bounding the model to narration and exploration while outsourcing determination to a deterministic engine—is a strong defensive mechanism against hallucination.
The introduction of the Adjudicated Query pattern successfully formalizes the requirements for auditability by making completeness an asserted invariant computed *before* any output is generated. This moves compliance checking from a statistical approximation (like RAG) to a verifiable accounting function, addressing the fundamental need for defensibility in litigation.
A key implication is the management of trust in the delivery layer. The text addresses the "untrusted renderer" risk by employing specific payload techniques—embedding caveats as un-strippable suffixes and supplying real aggregates—demonstrating that linguistic safeguards must be engineered structurally into the data flow, not merely suggested to the model. This suggests that building trustworthy AI interfaces requires an immutable separation between generative fluency and deterministic accountability, forcing users to recognize when they are engaging with a reasoning system versus a guaranteed record.
What follow-up questions remain: How does this pattern handle dynamic rule interpretation across overlapping jurisdictions where rule sets themselves conflict? If the "population" definition is contestable, how is the initial predicate established without introducing subjective judgment before the sweep begins? And if accountability is threaded through the finding (as suggested), what are the systemic costs of linking human identity to automated compliance findings, and does this introduce a new layer of privacy or liability risk?