---
type: "post"
id: "mathematical-proof-powered-by-ai-and-no-hallucinations-319bff"
slug: "mathematical-proof-powered-by-ai-and-no-hallucinations-319bff"
title: "Mathematical proof, powered by AI. And no hallucinations!"
description: "Amazon says it has made three tools in its AutoCorrode suite publicly available to connect AI agents with the Isabelle proof assistant."
retrieval_nugget: "Amazon says it has made three tools in its AutoCorrode suite publicly available to connect AI agents with the Isabelle proof assistant. The tools let agents generate and revise formal proofs against a machine checker, turning verification into feedback during work rather than a final review step."
published_at: "2026-11-21"
updated_at: "2026-11-21"
record_date: "2026-11-21"
date_kind: "published_at"
topics: ["claude"]
entities: ["Amazon","AutoCorrode","Isabelle","I/Q","I/R","Isabelle Assistant","Claude Opus 4.6","Model Context Protocol"]
related_patterns: ["scientific-agents-need-artifact-evidence"]
novelty_label: "new_fact"
what_changed: null
prior_refs: []
source_url: "https://www.amazon.science/blog/mathematical-proof-powered-by-ai-and-no-hallucinations"
source_urls: ["https://www.amazon.science/blog/mathematical-proof-powered-by-ai-and-no-hallucinations"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:mathematical-proof-powered-by-ai-and-no-hallucinations-319bff"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/mathematical-proof-powered-by-ai-and-no-hallucinations-319bff/","markdown":"https://newruntime.com/posts/mathematical-proof-powered-by-ai-and-no-hallucinations-319bff.md","json":"https://newruntime.com/posts/mathematical-proof-powered-by-ai-and-no-hallucinations-319bff.json"}
markdown_url: "https://newruntime.com/posts/mathematical-proof-powered-by-ai-and-no-hallucinations-319bff.md"
---

# Mathematical proof, powered by AI. And no hallucinations!

## Retrieval answer

Amazon says it has made three tools in its AutoCorrode suite publicly available to connect AI agents with the Isabelle proof assistant. The tools let agents generate and revise formal proofs against a machine checker, turning verification into feedback during work rather than a final review step.

## Summary
Amazon says it has made three tools in its AutoCorrode suite publicly available to connect AI agents with the Isabelle proof assistant. The tools let agents generate and revise formal proofs against a machine checker, turning verification into feedback during work rather than a final review step.

## What happened
Amazon Science’s November 18, 2026 announcement describes I/Q, I/R and Isabelle Assistant, tools that connect an AI agent to an Isabelle session through the Model Context Protocol (MCP). Isabelle is a proof assistant: it checks formal statements and the steps offered to establish them. The announcement frames this as a response to a familiar problem with language models: they can produce plausible but incorrect reasoning. In a formal system, the agent’s proposed proof can be checked by the assistant, and that check can be interleaved with generation.

The three tools offer different ways to arrange that interaction. I/Q lets an agent communicate with a live Isabelle session running on the user’s computer. The user can temporarily hand the agent control so it can fill in formal material, inspect context or offer advice. I/R lets an agent develop proofs or other formal material while the user keeps working in the proof assistant. The agent still communicates with Isabelle to learn whether its proofs work, but it does not take control of the user’s session in the same way. Amazon says I/R can also create a separate Isabelle subprocess that works in the background. Isabelle Assistant, powered by I/Q, provides a chat window in which an agent can observe and modify the user’s session; the announcement describes it as a convenient, out-of-the-box option that is less powerful than the full I/Q and I/R package.

Amazon’s examples use Claude Opus 4.6 with Isabelle Assistant. They include a proof that √2 is irrational, a result about blocks of consecutive composite numbers, and the definite integral of sin x from 0 to π. For the composite-number example, the agent noticed that a supplied informal argument established only N−1 consecutive integers and corrected the result. For the integral, the announcement says the agent generated a proof that includes the derivative of −cos x, as needed for the fundamental theorem of calculus. Amazon characterizes these as easy examples, while saying harder formalization tasks can benefit from I/Q and I/R.

The announcement presents a human-agent-assistant collaboration, not an autonomous mathematician. People can give hints, pause the agent to correct something, and make decisions on difficult problems. Amazon says its researchers use AI-powered proving in everyday AWS work, including lengthy tasks such as refactoring messy proof developments or studying traces to diagnose problems. It does not give a release date for each tool separately, a price, or a supported-version matrix in the described announcement. The tools are part of the AutoCorrode verification suite, and the connection described here is specifically to Isabelle.

## What the archive adds
New Runtime’s earlier coverage of Google DeepMind’s AlphaProof Nexus, on [October 15, 2026](https://newruntime.com/posts/google-deepmind-s-alphaproof-nexus-solved-9-open-07c5ac/), described formal proof search gaining research value when candidates are paired with machine-checkable proof obligations. Amazon’s announcement is a new event on that broader thread: it focuses on tools for connecting an agent to Isabelle during proof construction, rather than a result solving open problems.

## Trend signal
This reinforces the idea that scientific-agent progress is more useful when it produces artifacts others can inspect and check. Here, the artifact is a formal proof that Isabelle can validate, with checking interleaved with generation. The announcement does not establish that this approach scales across mathematical research, but it shows a concrete workflow for making some agent output machine-checkable. This aligns with [scientific-agents-need-artifact-evidence](https://newruntime.com/patterns/scientific-agents-need-artifact-evidence/).

## Connections to other events
New Runtime’s [November 14, 2026 coverage of OpenAI’s Yang–Mills claim](https://newruntime.com/posts/science-openai-still-withheld-the-1-147-page-3721ad/) highlighted the distinction between a formal check and a complete account of correctness. Amazon’s Isabelle workflow meets the same broader question from another direction: formal systems can check proof steps, while people still need to understand what has been formalized and what the result establishes.

## Why it matters
For teams building agents in domains where errors are costly, the useful design choice is to make verification part of the working loop. Isabelle can reject an invalid proof step while the agent is still composing a result, giving the model a concrete signal to revise rather than leaving a person to discover a flaw in a polished final answer. The tool split also exposes a practical trade-off: I/Q can work directly in a live session, while I/R leaves the user’s document alone and runs separately. That matters for collaborative work, where an agent’s access to the working state and the human’s ability to continue are both design decisions.

The examples are mathematical, but the principle extends only where an appropriate formal checker exists and the problem can be expressed in its language. A checker can establish that a proof satisfies Isabelle’s formal rules; it does not decide whether the chosen theorem answers the right question or whether a human-readable explanation is clear. For builders, the broader lesson is to pair generation with an external verifier and preserve a human role for framing and interpretation, rather than treating fluent output as evidence of correctness.

## Limits and uncertainty
Amazon’s examples demonstrate selected proofs, not a measured success rate across a benchmark or a range of research problems. The announcement does not quantify how often an agent completes a proof without human hints, how much time the tools save, or what compute and setup costs are involved. The examples use Claude Opus 4.6, so they do not establish comparable results for other models. Nor does machine checking alone show that a formalized theorem captures the intended informal claim or that the proof is easy for a mathematician to review. Amazon says harder problems still require human thought; the degree of autonomy and reliability in those cases remains unmeasured in the announcement.

## Reader takeaway
Amazon’s AutoCorrode tools make Isabelle available as an active verifier in an agent’s proof-writing loop. The important shift is not that a model can produce mathematical-looking text, but that an external proof assistant can check formal steps as they are produced. Where a trustworthy checker exists, that is a more useful foundation for agent work than relying on the model’s confidence.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Claude](https://newruntime.com/topics/claude.md): Explore the claude topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)
