BootLoops Lets AI Audit Its Own Calculations

Physicist Matthew Schwartz introduced BootLoops 1.0, an open-source software package, in a post published on Anthropic’s site on October 1. The software bundles together tools for high-precision scientific computation along with the procedures needed to verify the results they produce. Schwartz created and maintains the project; it is not an officially supported Anthropic product.

How does it check the AI’s math?

BootLoops is designed to let an AI run a calculation and then check its own work. Users download the code, connect it to an AI agent (a system that carries out tasks on its own), and hand it a problem. The AI reads the tool descriptions, proposes a computational plan, and then confirms the output using the verification routines that come with each tool.

During the calculation itself, the software handles symbolic expressions precisely and keeps track of error bounds. It also includes procedures for cross-checking a result against the same quantity computed by a different route. The public repository contains 49 tool packages. The design gives the AI access not just to methods for computing, but to methods for auditing what it computed.

30 integrals, and what came of them

Schwartz reported that Claude used BootLoops to evaluate 30 integrals over the course of several weeks. Of those, 15 reproduced results that were already known, while the other 15 were new calculations, he said.

Earlier experiments using Claude as a research assistant demanded a great deal of human labor. Schwartz explained that he had to intervene frequently to rewrite prose and steer the research back on course. BootLoops is not tied to any particular model, and it packages computational tools together with verification procedures. That makes its role different from general-purpose tools like Claude Code or Codex, which can be pointed at almost any task.

Even with verification routines in place, researchers still have to inspect the results themselves. Schwartz noted that an AI may declare a calculation finished prematurely, or get the math right and still misinterpret what it means. The repository states that the software is not suitable for clinical, actuarial, payment, regulatory, or public-safety decisions.

What you need to try it

The code is available from the public repository. The main code in the core repository is released under the MIT license, which lets anyone use and modify it free of charge. Some of the accompanying code carries other open-source licenses. To actually run it, though, you will need to supply your own access to whichever AI model you plan to use, along with your own compute. Certain features require installing external computation programs separately.

Researchers in South Korea would do well to check their installation environment in advance. The published testing environment is Linux-centric, and macOS was not part of the release testing. The core of BootLoops lies in attaching both computational tools and verification procedures to an AI. The open questions going forward are whether the same results hold up on other research problems, and how much of the burden of reviewing both the math and its interpretation the system can actually lift from human shoulders.