AWS Details Adjudicated Query Pattern for Lease Compliance Sweeps in Amazon Quick

The problem: checking every lease and proving it

AWS frames the issue around a portfolio operator that holds 50,000 leases across multiple states. Each state publishes landlord-tenant statutes, such as late-fee caps, notice periods and security-deposit limits, and these change on the legislature's schedule. When a rule changes, the compliance team must work out which leases are now out of line.

At small volume a paralegal can read the leases, and the answer is trusted because a human stands behind it. Past a certain size the work moves to software, and the result becomes a number on a screen that nobody can independently verify. AWS names two properties that follow from this:

• Provable completeness: a claim such as "we checked all 22,910 Texas leases" must be true and demonstrable. A record that was never assessed must be reported as unevaluated rather than silently omitted.
• Defensibility: a finding may be challenged months later in litigation, an audit or a regulatory examination. Defending it means knowing which version of which rule was applied, to which clause text, by what method, on what date, and by whom.

AWS argues that Retrieval Augmented Generation (RAG) cannot satisfy either property, because similarity search has no threshold that means all records and a ranked sample never knows what it excluded. Text-to-SQL carries its own risk: a hallucinated predicate can silently shrink the population, and the resulting number looks exact even when the scope is wrong.

How the pattern works

Adjudicated Query is described as a bounded conversational layer over a deterministic rules engine. The model does two things only: it translates a natural-language question into a call on a fixed set of typed operations, and it narrates the result. It never writes a query, never fixes the population and never performs a determination.

Rules are stored as versioned data, not code. The engine knows generic comparison operators (gte, lte, equals, exists) and contains no branch naming a jurisdiction or topic. As a result, a law change is a rulebook row edit rather than a code deployment.

Each sweep produces a completeness receipt, an asserted invariant in which compliant plus in-breach plus ambiguous plus unreadable must equal scanned. It is computed from counts and checked before anything is saved, so a run that cannot account for its population never finishes. The chat surface shows counts, the receipt and a labeled sample, while the full result set lives on a dashboard reading the same data store.

AWS compares the approaches this way: RAG gives a ranked sample and makes completeness structurally impossible, text-to-SQL gives a claimed but unprovable population, while a rules engine with BI and Adjudicated Query both give an exact and proven population. The new pattern adds natural-language access to the rules engine plus BI approach without giving up the guarantee.

Reference architecture and the six tools

A compliance officer uses two surfaces in Amazon Quick: a chat agent for questions and an Amazon Quick Sight dashboard for browsing the full results. The chat agent gets an OAuth token from Amazon Cognito, then sends Model Context Protocol (MCP) requests over an Amazon API Gateway HTTP API, which validates the token before passing requests to an AWS Lambda function. That function hosts the MCP server and the rules engine and uses Amazon Aurora Serverless v2 (Postgres with pgvector) through the RDS Data API. The dashboard reads the same Aurora store through a VPC connection, which makes the completeness receipt a single source of truth.

The MCP server exposes exactly six tools:

• sweep_compliance: an exhaustive sweep against the rules in force on a stated date. Results are official and written as findings.
• simulate_rule_change: tests one rule at a proposed value against the approved baseline. Exploratory, directional counts only, nothing recorded.
• explore_clauses: top K by semantic similarity within a filtered population. Interpretive, and it cannot answer "how many".
• get_finding: the complete evidence chain for one finding.
• list_rules: the rulebook in force on a date, with versions, citations and approvers.
• check_connection: a liveness check that touches no data.

Bounded operations translate to set-based SQL with rule values bound as parameters, so no natural language reaches the query layer. Amazon Bedrock is called only on the exploratory clause-search path, using Amazon Titan Text Embeddings V2 for similarity ranking and Anthropic Claude Sonnet 5, invoked through a cross-region inference profile, for qualitative clause assessment. According to AWS, no model is involved in compliance sweeps. AWS also notes that Bedrock model availability varies by Region and should be confirmed before deployment.

What it means for compliance teams

AWS positions the pattern for cases where accountable users need conversational access and a missed record is a liability rather than a mild inconvenience. Because the model can only choose from operations whose population logic was written, reviewed and tested by people, it cannot compose a wrong population.

Although the post uses lease compliance as its example, AWS says the pattern also applies to other high-stakes domains, including sanctions screening, insurance claims adjudication and export control. The post also includes a working sample that readers can deploy and run end to end.