36 manuscripts across 18 fields from the BootLoops harness run with Claude
Matthew Schwartz's open-source BootLoops harness, operated by Claude, produced 36 manuscripts with 19 coauthors between July and September 2026, mostly exact or certified computations carried from physics and computer science into fields such as population genetics, phylogenetics, ecology and health policy; eight are still in preparation and none is yet listed as published in a journal.
- Lab
- Independent
- Model
- Claude (the essay names Claude Fable 5 for the initial porting; versions per result are not recorded)
- Field
- Physics
- Date
- 2026-10-01
- Human collaborators
- Matthew D. Schwartz
What was found
BootLoops pools programs from particle physics, experimental mathematics, computer algebra and numerical analysis in one toolkit that a language-model agent can operate and extend. The overview paper (arXiv:2610.02286) says the computations were run by Claude with a human setting the targets and the standards, and that the package itself was written by AI apart from forks of public codes. The essay on Anthropic's blog counts 36 manuscripts in 18 fields with 19 coauthors over three months, chosen from about 400 candidate problems; the project's manuscript index lists the overview and 35 papers in nine groups, eight marked in preparation, and says every manuscript has been read by humans for accuracy, that many were drafted first by language models, and that those not yet fully checked are watermarked preliminary. The recurring move is to replace an estimate by an exact or certified answer: Feynman integrals in closed form by bootstrap, Bayesian evidence for four- and five-species phylogenetic trees as exact rational numbers, the site-frequency spectrum of a selected allele at any selection strength, Medicare star-rating cut points at the true optimum of their clustering criterion, and exact operating characteristics of published adaptive trial designs. Coauthors include Scott Edwards and Paul Lewis on phylogenetics, Michael Desai on population genetics, James O'Dwyer on ecology and Isaiah Andrews on economics. This registry's 2026-10-01-anisotropic-watson-integral is one result from the project.
Novelty check
Read the BootLoops manuscript index, the overview paper's model and methods sections and the Anthropic essay on 2026-10-16. The registry carried one result from the project, 2026-10-01-anisotropic-watson-integral, and nothing else citing bootloops.ai or the overview. Novelty was not checked manuscript by manuscript: each needs its own check against its field's literature before it is entered, and several recompute published or regulatory numbers rather than claim new theory, which changes what such a check means for them. Science's article on the project could not be read from here.
Caveats and known objections
A container for the release, graded claimed: the essay and the index are the authors' own account, nothing is refereed, eight manuscripts are not public, and the index itself says some drafts have not been fully checked by humans. The strength of the work varies by kind, which is why each manuscript needs separate grading: exact rational evidences, interval-certified integrals and proven enumerations can be rerun by anyone with the published code, while the ecology fits and policy reanalyses rest on modelling choices. The overview names the failure mode its methods guard against, a bug in a computer-assisted search that silently skips part of the space, and uses planted synthetic targets as controls. Autonomy is graded ai-assisted, the weakest reading that fits the whole release: the overview says Claude ran the computations and wrote the package with a human setting targets and standards, while the essay warns that the automated checks still cannot be trusted and that Claude's conclusions can be wrong, and says the human supplies the taste in choosing problems; how much each manuscript owes to the model is not documented per paper. Schwartz is a visiting researcher at Anthropic and the essay ran on Anthropic's blog, but it states that BootLoops is not an Anthropic project and is owned and maintained by Schwartz. Which model versions produced which results is not recorded.
Nobody outside the lab has checked this yet.
Reading the primary source closely enough to say whether it supports the claim counts as a check, and you are credited on the entry.
Or on GitHub: submit a check challenge the grade send a correction or send a pull request
Flag this for triage
Signals order the review queue and nothing else. They are never published, and they never move a grade: that takes a citation.
Entry history (1 event)
- AddedEntered the registry graded Claimed and AI-assisted.
Entries are never deleted. A grade that does not hold up is downgraded on the record, with the reason beside it.
Graded claimed for verification and ai-assisted for autonomy. What these mean.
Cite this entry
whataifound.org. (2026). 36 manuscripts across 18 fields from the BootLoops harness run with Claude. whataifound.org: A Registry of AI Scientific and Mathematical Discoveries. https://whataifound.org/finding/2026-10-01-bootloops-manuscriptsBibTeX
@misc{whataifound-independent-2026-manuscripts,
title = {36 manuscripts across 18 fields from the BootLoops harness run with Claude},
author = {{whataifound.org}},
year = {2026},
howpublished = {whataifound.org: A Registry of AI Scientific and Mathematical Discoveries},
note = {Result by Independent. Verification: Claimed. Autonomy: AI-assisted.},
url = {https://whataifound.org/finding/2026-10-01-bootloops-manuscripts}
}