research

AWS releases I/Q, I/R and Isabelle Assistant to connect AI agents to the Isabelle proof checker via MCP

TL;DR

AWS has made three tools in its AutoCorrode verification suite publicly available: I/Q, I/R and Isabelle Assistant. They connect AI agents to the Isabelle proof assistant through the Model Context Protocol, so generated proofs are checked by computer. Amazon Science authors demonstrate the setup using Claude Opus 4.6.

3 min read
0

AWS has released three tools in its AutoCorrode verification suite that let AI agents write and verify formal mathematical proofs in the Isabelle proof assistant: I/Q, I/R and Isabelle Assistant. All three are publicly available, according to an Amazon Science blog post by Hanno Becker and Larry Paulson published November 18, 2026.

How it works

The tools connect an agent to Isabelle through the Model Context Protocol (MCP), which the authors say is supported by all the leading AI agents. The authors' premise is that AI models can hallucinate, so a computer should check their proofs, with checking interleaved with generation to keep the agent on track. To be checked, a proof must be written in a formal language, and proof assistants are applications built for writing and checking such proofs.

The authors name Lean, HOL Light and Isabelle as examples. According to the post, AWS uses HOL Light to formally verify high-performance cryptographic code, including for post-quantum cryptography, and uses Isabelle to verify the Nitro Isolation Engine. They write that most proof assistants suit this task, and the post focuses on Isabelle.

The authors say some proof assistants, such as HOL Light, offer a simple command-line REPL that is easy for an agent to use. Others, such as Lean with its Microsoft VS Code extension, have sophisticated interfaces built for humans. An agent needs something simpler and more direct.

The three tools

  • I/Q: Lets an agent communicate with a live Isabelle session on the user's computer. The user can temporarily hand control of the proof assistant to the AI to fill in formal material or to assess context and give advice.
  • I/R: Lets an agent develop proofs or other formal material while the user keeps working in the proof assistant. The agent still talks directly to Isabelle to learn whether its proofs work, but it does not take control as I/Q does.
  • Isabelle Assistant: Built on I/Q, it gives an Isabelle user a chat window to an AI agent that can observe and modify the session, developing or fixing proofs. The authors say it is less powerful than the full I/Q and I/R package but runs in a convenient window out of the box.

Demonstrations with Claude Opus 4.6

The post shows Claude Opus 4.6 producing Isabelle proofs through Isabelle Assistant for three classic results.

  • Irrationality of the square root of two: The authors call the formalized proof "surprisingly idiomatic, though more detailed than strictly necessary."
  • N consecutive composite numbers: The authors supplied a flawed proof based on N! + 2 through N! + N. According to the post, Claude noticed that this argument yields only N - 1 consecutive integers and made the correction. It produced three facts, the third being the theorem in correct form.
  • Integral of sin x over [0, π]: The integral equals two. The authors asked only for "a nice simple example involving calculus," and say a plain-English description of the integral would also have worked.

The authors frame the result as AI assisting rather than replacing mathematicians, in a three-way collaboration between the mathematician, the AI agent and the proof assistant.

What this means

The release addresses a specific problem: AI-generated proofs can contain errors, and a proof checker rejects any step that does not hold. Because I/Q hands over session control and I/R runs alongside the user, mathematicians and verification engineers can choose how much autonomy to give the agent. The three demonstrations are textbook-level results, so the post shows the workflow rather than performance on research-level problems. It gives no benchmark scores or pricing.

Comments

Loading...