News · 2026-10-02

Matthew Schwartz reports a Claude-assisted science pipeline built around checkable calculations

Harvard physicist Matthew Schwartz reports that Claude and his BootLoops toolkit accelerated exact calculations and supported a three-month pipeline of 36 manuscripts across 18 fields. His October 1 account also says domain experts were essential for turning technically correct results into scientifically meaningful work, and the pipeline is not a set of 36 peer-reviewed discoveries.

Key facts

• Schwartz published his guest account on Anthropic’s science blog October 1, 2026.
• He reports 30 completed Feynman-integral calculations: 15 reproductions and 15 previously uncomputed results.
• The larger pipeline came from roughly 400 candidate problems.
• Primary sources are Schwartz’s account and the BootLoops manuscript index.

The compelling part of this story is the feedback between calculation, software, and scientific judgment. Difficult mathematical work can occupy a research group for a long time. A model that helps port algorithms, write missing code, and recognize a reusable method can speed that work, even when it cannot decide which result will matter to a discipline.

BootLoops is an open-source toolkit for exact calculations in quantitative science. The underlying semi-numerical approach combines highly precise numerical sampling with analytic constraints. Its purpose is to narrow a large set of possible expressions until numerical information can identify the remaining coefficients. This does not turn an arbitrary decimal approximation into proof of any formula the model proposes.

Imagine knowing that a tune must use a particular set of notes and rhythms. Listening to a few extremely clear fragments can distinguish among the remaining candidate scores. The restrictions make the identification possible; clearer recordings alone would not recover every imaginable composition. Scientific constraints play the corresponding role in these calculations.

The cited 2025 preprint, co-authored by Schwartz, concerns analytic regression of Feynman integrals from high-precision numerical sampling. That provides methodological background rather than making today’s guest post a new paper release. The news is a researcher’s report about an AI-assisted workflow and the broader research program it supported.

A language model can contribute by translating implementations, assembling a common software interface, and applying a known technique where a different field uses similar mathematics. The BootLoops project site provides an inspectable manuscript inventory. Claims of speed or novelty still need to be assessed at the level of the particular calculation or manuscript, rather than inferred from the size of that inventory.

Schwartz describes one reproduction taking about 20 minutes, then the set of 30 integrals. The division between 15 known results and 15 he says were previously uncomputed matters. Reproduction tests whether the workflow can recover established work. A new result raises additional questions about prior literature, mathematical validity, assumptions, and scientific usefulness. Both are valuable, but they are different evidence.

The account’s reception is unusually informative because it includes collaborators’ skepticism. James O’Dwyer and Michael Desai were impressed by technical accomplishments without initially finding the science compelling. Their guidance helped redirect subsequent work. That is a stronger corrective than a generic statement that a human was somewhere in the loop: the humans changed what the project should pursue.

Schwartz’s concise instruction is “Supply the taste.” The phrase, checked against the primary post for this synthesis, captures the missing judgment. An equation can be evaluated correctly while answering an unimportant question. Our lessons on agent harnesses and program synthesis help explain how a toolkit can support execution without supplying all the scientific criteria for success.

The strongest counter-argument is that spectacular productivity figures can blur preliminary outputs with established contributions. A manuscript is not automatically a discovery, and peer review is not the only missing step: independent replication and domain significance also matter. The manuscript index itself labels some projects as preliminary or still in preparation, making the caveat part of the available record.

There is also an affiliation disclosure. Schwartz was a visiting Anthropic researcher during the project. That does not invalidate the work, but readers should know the report appears on the model provider’s site and comes from an affiliated participant. The dossier establishes an author-reported program, not an independent audit of all its outcomes.

The practical lesson is to seek work with strong checks, reusable tools, and experts who can question relevance. The report gives a concrete example of AI widening the reach of technical methods across fields. It does not establish autonomous scientific discovery or the disappearance of human conceptual work. The next decisive evidence lies in the individual artifacts: reproducible calculations, verified new results, and scientific questions that experts judge worth answering.

Primary source, verified: read the paper →

Ground Truth. (2026, October 2). Matthew Schwartz reports a Claude-assisted science pipeline built around checkable calculations. Ground Truth. https://groundtruth.day/news/bootloops-claude-shaped-science-human-judgment.html

@misc{groundtruth:bootloops-claude-shaped-science-human-judgment,
  title  = {Matthew Schwartz reports a Claude-assisted science pipeline built around checkable calculations},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {oct},
  url    = {https://groundtruth.day/news/bootloops-claude-shaped-science-human-judgment.html}
}

Topics: science · agents · scientific-computing · anthropic · evaluation