Hi - I answer from the OpenSmartRoute documentation: routing, the API, plans and quotas, self-hosting. Ask away, or open a support ticket if you need a person.
Grounded in the docs - follow a source before acting on it.
AWS introduces Adjudicated Query pattern for lease compliance checks
AWS's new pattern combines generative AI and deterministic rules to verify thousands of leases for compliance, ensuring transparency and accountability.
Key points
Pattern applies to high-stakes compliance like leases, sanctions, and insurance.
Uses Amazon Quick chat interface with a rules engine for deterministic decisions.
Supports checking 50,000 leases across multiple states with proof of completeness.
Deploys on AWS with Aurora Serverless v2, Lambda, API Gateway, and Bedrock models.
Why it matters: Provides a reliable, auditable way to verify large datasets against changing regulations, reducing risk and increasing trust.
By OpenSmartRoute editorial · written through the router by llm-small
From AWS machine learning blog - “Sweep thousands of leases for compliance using Amazon Quick and the Adjudicated Query pattern”
Architecture diagram: Amazon Quick chat agent and Amazon Quick Sight dashboard both read from an Aurora store, with a Lambda-hosted MCP server and rules engine mediating chat requests through API Gate. Image: AWS machine learning blog (original)
Introduction - Announcing the new Adjudicated Query pattern for lease compliance
AWS has introduced a new pattern called Adjudicated Query to help verify lease compliance. This pattern combines generative AI with a deterministic rules engine. It allows users to ask compliance questions in a chat interface while keeping the actual decisions in a strict, rule-based system. This approach makes it possible to check thousands of leases efficiently and reliably. It also provides proof that every lease was reviewed, which is important for accountability.
The pattern is designed for high-stakes compliance tasks. It is especially useful when verifying many leases against changing laws. The pattern can be applied to other areas like sanctions screening, insurance claims, and export control. The goal is to make compliance checks transparent, complete, and defensible. It ensures that no lease is missed and that decisions can be explained later if needed. A working sample implementation is available for deployment and testing.
The compliance challenge - Verifying thousands of leases against changing laws
Checking large numbers of leases for compliance is difficult. Lease portfolios can contain tens of thousands of records. Laws and regulations in different states change frequently and unpredictably. Compliance teams must stay updated on these changes and verify each lease accordingly. When laws change, they need to identify which leases are now out of compliance.
Traditionally, humans review leases manually. This process is accurate but slow and expensive. As the number of leases grows, manual review becomes impossible. Software tools can help, but they often lack transparency. They may produce results that are hard to verify or defend later. This creates a risk of silent omissions, where some leases are not checked at all.
The main challenge is to verify every lease thoroughly and prove it. Teams need to demonstrate that they checked all records, not just a sample. They also need to be able to explain how they made each compliance decision. This requires a system that is both complete and defensible, which is hard to achieve with existing AI methods.
Google launched EmbeddingGemma 2, a compact open model that handles text, code, images, video, and audio. It uses a single shared vector space to enable unified search across all media types.
Google launched EmbeddingGemma 2, a compact open-weight model that maps text, images, and audio into one vector space. It runs locally on phones with minimal RAM and enables instant on-device semantic search.
Properties of the problem - Ensuring provable completeness and defensibility
Two key properties are essential for lease compliance verification. The first is provable completeness. This means the system must confirm that every lease was checked. If some leases are not assessed, the system must report them as unevaluated. It cannot silently skip any record. This property ensures that the compliance report is trustworthy and comprehensive.
The second property is defensibility. This means that each compliance decision can be justified later. It must be possible to trace which rule was applied, what clause was checked, and who made the decision. This is important for audits, legal challenges, or regulatory reviews. The system must store detailed records of every step taken during the compliance check.
Achieving both properties is difficult. Many AI methods fall short. Retrieval-augmented generation (RAG) can help with accessibility but does not guarantee completeness or defensibility. Similarity search ranks records but does not ensure all are included. Text-to-SQL can narrow the gap but risks silently excluding records if the query is incorrect. These limitations make them unsuitable for high-stakes compliance.
Limitations of existing AI methods - Why RAG, similarity search, and Text-to-SQL are insufficient
Retrieval-augmented generation (RAG) is a technique that combines retrieval of relevant documents with language generation. It helps users ask questions and get answers based on large data sets. However, RAG cannot guarantee that all relevant records are retrieved. It may miss some, leading to incomplete checks.
Similarity search ranks records based on how closely they match a query. It can help find similar leases but does not set a clear threshold for inclusion. It cannot confirm that all leases have been checked. This makes it unsuitable for compliance where completeness is critical.
Text-to-SQL converts natural language questions into SQL queries to fetch data from databases. It narrows the gap between human questions and structured data. But it carries a risk: a hallucinated predicate can silently exclude some records. The resulting number may look exact but could be based on an incomplete set. This makes it unsafe for high-stakes compliance checks.
These methods are useful for exploration or informal analysis. But they cannot provide the guarantees needed for official compliance reports. They lack the ability to prove that every lease was reviewed and that decisions are fully traceable.
The Adjudicated Query pattern - Combining rules and conversational AI
The Adjudicated Query pattern is a new approach that combines a bounded, rules-based system with conversational AI. It creates a layer where users can ask questions naturally, but the actual data processing remains deterministic and transparent. The AI component translates natural language questions into a fixed set of operations. It never writes new queries or modifies the population logic.
Behind this layer is a rules engine that is versioned and data-driven. Rules are stored as data, not code. They are simple comparison operations like greater than, less than, or equal. When laws change, the rules are updated by editing data rows, not deploying new code. This makes the system flexible and easy to audit.
Every compliance sweep produces a completeness receipt. This is a report showing how many leases were checked, how many are compliant, and how many are ambiguous or unreadable. The system asserts this report before saving results. If it cannot account for all leases, it does not finish the sweep. This guarantees that no lease is silently skipped. The separation of the conversational interface and the rules engine ensures transparency and accountability.
System architecture - Components and their roles
The system architecture includes several AWS services working together. Amazon Aurora Serverless v2 hosts the rulebook, lease data, and results. It acts as a single source of truth for all information. This relational database allows easy counting and verification of completeness.
AWS Lambda hosts the MCP (Model Context Protocol) server and the rules engine. It translates user questions into fixed SQL operations and manages interactions with the database. The MCP server exposes six tools, each with a specific purpose, such as sweeping all leases, testing individual rules, or providing summaries. These tools are designed to prevent any accidental omission or incorrect population.
Amazon API Gateway provides a secure front door. It uses JSON Web Tokens (JWT) issued by Amazon Cognito for authentication. The chat interface communicates with the MCP server through this API. Amazon Bedrock models are used only for exploratory clause searches, not for official compliance decisions. They assist in semantic similarity ranking and clause assessment but do not influence the final outcome.
The Amazon Quick Sight dashboard connects directly to Aurora. It displays the full results set, allowing compliance officers to browse and verify each record. Both the chat interface and the dashboard read from the same database, ensuring consistency. This architecture guarantees that the completeness receipt reflects the actual data checked.
Operational details - Enforcing guarantees and preventing silent omissions
The system enforces guarantees through strict control of data flow and operations. The rules engine only executes predefined, versioned operations. It does not generate queries dynamically or modify data outside approved procedures. This prevents the risk of silently excluding leases or changing the population without traceability.
The completeness receipt is computed before any results are stored. It sums counts of compliant, breaching, ambiguous, and unreadable leases. If the total does not match the number of scanned leases, the process halts. This ensures that every lease is accounted for and no silent omission occurs.
The separation of the conversational interface from the deterministic engine is crucial. The chat interface only narrates results and counts, never performing the actual compliance logic. The full set of results is stored in Aurora and accessible through the dashboard. This design maintains transparency and makes it easy to verify the process later.
Safeguards against model paraphrasing - Preventing AI from misrepresenting data
One challenge is preventing the AI from paraphrasing or misrepresenting data in summaries. AI models can sometimes strip caveats or infer information that is not present in the original data. To address this, safeguards are built into the system.
The data provided to the model includes explicit labels and repeated mode indicators at multiple levels. These labels help ensure that the model presents accurate information. Caveats are phrased as un-removable brackets, not leading labels, so they survive paraphrasing attempts.
Real aggregates are computed over every record and provided directly to the model. This prevents the model from guessing or inferring totals based on samples. The design principle is that any safeguard in the payload must survive paraphrasing to be effective. This approach helps maintain the integrity of the compliance report.
Implementation steps - Deploying the sample stack and setting up
Implementing the system involves several steps. First, an AWS account with access to Amazon Bedrock models is needed. The account should be in the US East (N. Virginia) Region, where the models are available. Credentials for AWS CLI must be configured correctly.
Next, clone the sample repository from GitHub. Set up a Python environment and install required packages. Bootstrap the AWS Cloud Development Kit (CDK) and deploy the stack. The deployment creates the Aurora database, Lambda functions, API Gateway, and other components.
After deployment, run migration scripts to set up the database schema and rulebook. Generate a deterministic corpus of lease data and ingest it into the system. Then, set up the Amazon Quick Sight dashboard to visualize results. Finally, run acceptance tests to verify the system works as intended.
The entire process is designed to produce a reproducible environment. The data and results are deterministic, making it easy to verify and audit compliance checks. The sample implementation provides a foundation for building custom solutions tailored to specific compliance needs.
AWS released guidance on using ISO/IEC 42005:2025 for AI impact assessments. The company offers tools to help customers integrate these checks into their risk management.
