Claude Pushes an Amplitude Calculation from Eight Loops to Nine, Anthropic Says

Anthropic has reported a nine-loop calculation performed by Claude. The previous published result for this amplitude ended at eight loops. That is one more order of the perturbative expansion, and each successive order costs noticeably more than the one before. The text available to us contains no details of the calculation itself. There is, however, a long analysis on Habr, where the author turned the news into a practical experiment in controlling scientific code written by AI.

What "nine loops" means

In quantum field theory, prediction accuracy is increased order by order in perturbation theory. One order corresponds to one loop in Feynman diagrams. Each new loop increases the number of diagrams, the dimensionality of the integrals and the volume of algebra. In terms of labor, the step from eight loops to nine is incomparable to the step from one to two.

The task is hard for both humans and code. An error in a sign, a symmetry factor or the cancellation of divergences can easily hide among thousands of terms. So the statement "the model calculated nine loops" says nothing about the reliability of the result until a verification scheme is described. The Habr write-up contains neither the calculation itself nor its verification, so independent confirmation will have to wait for the publication.

A graphene thesis as a testing ground

The author of the analysis did not simply retell the press release. Using AI, he reconstructed the calculations from his own thesis on graphene and found an error that remained after his defense. This is the most valuable episode in the piece: here the AI acted as a reviewer of old human work.

From this episode the author moves on to a method he calls process decomposition. The calculation is split into stages, and each stage has:

• inputs and outputs;
• contracts: what exactly the stage is required to guarantee;
• independent checks of the result by another route;
• checks of the checks themselves;
• deliberately introduced defects, to confirm that the controls catch them.

The last item is close to mutation testing in software engineering. If a check fails to notice a known error, its "green" status cannot be trusted.

What is inside the article

• Claimed result: an amplitude calculation at 9 loops versus 8 in the earlier publication.
• Scale: 95 illustrations, some of them animated.
• Code: complete Python code and 7 reproducible experiments.
• Line of exposition: from the Schrödinger equation and a single loop to bootstrap methods and symbols.
• Concept: "quantum pictorialism" as a possible foundation for scientific code in quantum theory.
• Verified fact: an error was found in the thesis, which had survived the defense.

The author presents the last point as a hypothesis, not as an established standard. The diagrammatic language is proposed as a structure in which code is easier to read and verify.

How to adopt the approach

If you hand calculations, simulations or analytical code to an AI, start with three steps:

1. Break the task into stages and write down a contract for each one: permissible inputs, the form of outputs, invariants. 2. For each stage, set up at least one check by a different route: a known limiting case, a numerical calculation against an analytical one, a comparison with a published value. 3. Once per cycle, plant a known defect in the code and see whether the tests catch it. If they don't, the checks need to be rewritten.

And one more point: judge a claim about "N loops" or any other record by whether the code, the contracts and the methods of independent verification have been published. Without them, it is so far a report of a result, not the result itself.