While AI hallucinations have been a major roadblock for enterprise adoption, automated reasoning offers a mathematical approach to verify AI outputs and prevent factual errors. Think of it as a rigorous fact-checker that uses proof-based verification instead of best guesses. Now available in preview on Amazon Bedrock!

How It Works in Three Steps

1. Analyzes program control structure
2. Verifies eventual truths within the program
3. Confirms invariant properties (what’s always true)

A Simple Example

For a program that classifies numbers as even or odd, automated reasoning:

• Maps out the logic flow (if divisible by 2, print “even”)
• Confirms it will eventually give an answer
• Proves it will always output either “even” or “odd” — nothing else

You might already be using automated reasoning without knowing it! Tools like IAM Access Analyzer and VPC Reachability Analyzer leverage these same principles.

Implementation on Bedrock

Bedrock makes implementation surprisingly simple. Upload your document and Bedrock will automatically create an automated reasoning policy. You then attach the policy to a guardrail and you’re ready to go.

You can even combine it with other techniques such as Retrieval Augmented Generation (RAG) for an extra layer of accuracy — RAG provides context at input, while automated reasoning validates the output.