What was released
Anthropic published a guest post on Wednesday by Matthew Schwartz, a theoretical physicist at Harvard and a visiting researcher at the company, describing BootLoops — an open-source toolkit he built for performing exact calculations in quantitative science. The code is on GitHub.
Schwartz describes it in plain terms: “BootLoops functions as a kind of harness for the LLM, much like Claude Code or Claude Science is a harness for Claude.” It is deliberately model-agnostic, and he says it can be driven by Claude, Gemini or ChatGPT.
The tally he reports is the part that has travelled: over three months, he and 19 co-authors produced 36 manuscripts across 18 fields, drawn from a pool of roughly 400 candidate problems.
The standard that makes the number mean something
A raw count of preprints is not evidence of anything. What makes this one worth reading is the acceptance test Schwartz set. A calculation is not finished when the model says it is finished; it is finished when the full functional form is known and a Python script can evaluate it to arbitrary precision on a laptop.
That is a verifiable bar, and it rules out the failure mode these systems are worst at. Schwartz is explicit about it: “Claude likes to declare victory too early.”
He started in his own field, with scattering amplitudes and elliptic integrals. Within weeks the system had computed 30 integrals — fifteen reproducing known results, fifteen computed for the first time — and generalised the machinery it had built for one class of integral to others, writing most of the software itself or borrowing it from mathematics rather than physics.
Where it went next
The work spread across ecology, population genetics, phylogenetics, economics, linguistics, genomics, cosmology and mathematical physics. According to The Decoder’s account of the post, that included solving a twenty-year-old equation in neutral biodiversity theory — with a finding that tree species composition changes about 4.5 times faster than the theory predicted — along with an analysis of 5.7 billion mutation pairs from the 1000 Genomes Project, an evaluation of 4,452 economics replication packages, and a word-stress database covering 6,072 languages.
The common thread is what Schwartz calls “Claude-shaped problems”: questions that are exactly stated, verifiable by computation, and bounded — the shape current systems handle well, rather than the shape science usually arrives in.
The caveat the author puts up front
Schwartz does not claim the expert has been removed from the loop. “Claude is technically correct, but the result was not all that interesting until the expert helped steer us,” he writes. Technical correctness and scientific significance are different properties, and only one of them is automatable here.
He is blunter still about the limit: “Most scientific progress comes from real-world data that has to be acquired, understood, and checked, with each round informing the next question.” A harness that closes exact calculations does not touch that.
What to watch
The manuscripts are preprints, and the test of this is peer review rather than the count. The more immediate question is whether BootLoops gets used by people who did not build it. An open harness that only its author can drive is a demonstration; one that produces checkable results in other people’s hands is a tool.
