Claude-shaped science

A guest essay by the physicist Matthew Schwartz on BootLoops, an open-source harness of scientific programs that Claude (Fable 5) built and operated under his supervision; he reports thirty Feynman integrals in closed form, fifteen of them new, and a run of projects in other fields that domain experts reshaped, and argues that current models do best on problems a known method from another field would solve. The results are in a companion arXiv overview, an NBER working paper and self-hosted manuscripts, most not yet peer reviewed.

Matthew Schwartz (guest post) · 2026-10-01 · original

Where can current language models add to science, if not by working like a human scientist? Matthew Schwartz, the physicist whose earlier guest post Vibe physics: The AI grad student described supervising Claude Opus 4.5 through one calculation, returns in this essay, published on Anthropic's Science Blog on 1 October 2026, with a different answer. Instead of pushing Claude toward deep conceptual problems, he looked for "Claude-shaped" problems: ones where a technique from mathematics, physics or computer science would settle a question in another field if anyone there knew it existed. The vehicle is BootLoops, a toolkit of programs that Claude ported, wrote and ran, first for scattering amplitudes and then for ecology, genetics, economics, linguistics and other fields.

The post is a first-person essay with no measurements of its own. Its technical companion is Schwartz's overview paper, BootLoops: an LLM-driven toolkit for exact quantitative science (arXiv:2610.02286v1, submitted 1 October 2026, 27 pages), which this note treats as the version of record for the toolkit and the integral results. The overview states results only for the integrals and for the Watson integral; for its other applications it describes the computation "without reporting results" (Table 2 caption). The individual results are in manuscripts listed on the project's website (bootloops.ai, "The manuscripts" page), in an NBER working paper and in a bioRxiv preprint; the code is on GitHub (the post links commit 66b680c of 1 October 2026; later commits in October added credits, notices and license headers for code derived from other authors' programs). Numbers below are from the post unless another document is named.

Disclosures. The post states that Schwartz "has been working as a visiting researcher at Anthropic" during the project, and that BootLoops is not an Anthropic project but is owned and maintained by him (section "Disclosure"). The arXiv overview, the bioRxiv forest preprint and the BootLoops website add that funding for the project was provided by Anthropic. The NBER paper's disclosure says that for that project Schwartz "worked as a contractor for Anthropic."

The argument

Schwartz describes an "impedance mismatch" between what scientists want and what language models do well (introduction): models need much guidance to produce anything of scientific value, and the headline results so far have mostly been "well-scoped applications of existing techniques." He proposes treating Claude as "the collaborator it actually is" (section "Claude, take the wheel!"), with broad knowledge, strong coding and fast reading of papers and data, but not, in his judgment, able to help with deep conceptual questions. His closing image is the convex hull (section "Outlook"): the region between fields, which humans could reach but no particular human has, is where he argues a harness like BootLoops is useful. The arXiv overview states the same thesis (Section 7): the approach "does not attempt to push out the deepest boundaries of human knowledge in one direction" but combines knowledge that already exists in different fields.

What BootLoops is

BootLoops is a harness, a layer of software through which a model is operated, open source and independent of the model that runs it (introduction). The arXiv overview's abstract names its main methods: integral reduction, differential equations and high-precision evaluation from particle physics, integer-relation fitting, exact enumeration, and ball (interval) arithmetic; Table 1 groups the principal programs in nine classes. Some are public codes, adopted or ported; for the Feynman-integral work alone, roughly one hundred tools were written where no public equivalent was found (Section 2.2). The paper states that "No single ingredient of the methods in BootLoops is new" (Section 1); its contribution is pooling them as working programs that a model can operate and extend.

The paper's description of the division of labor (Section 2.1) is that Claude runs as an agent, plans each calculation and writes new routines when tools are missing, "with a human setting the targets and the standards." It adds that the programs compute every number and "At no step is the model asked for a digit."

The amplitudes work

The project began with the semi-numerical S-matrix bootstrap for scattering amplitudes, which fixes an answer by physical constraints plus a few values computed to very high precision (section "Claude, take the wheel!"). Schwartz chose it because it needs coding across scattered packages and is checkable against the integral to as many digits as one likes. His first assignment for Claude Fable 5, released in summer 2026, was to port the field's code to a common framework. He reports that Claude reproduced the results of one of his papers in about 20 minutes, against weeks for his own code, and pointed out a more efficient algorithm he had not known. Asked to find unsolved amplitudes, Claude first limited itself to logarithmic function classes; at his prompting it moved to elliptic functions. After "only a few weeks," the post says, they had 30 integrals computed end to end: 15 reproductions of known results and 15 never computed before. The post's sequence suggests these are elliptic; the overview (Section 3.4) says the thirty run from one to four loops and range over polylogarithmic, elliptic, K3 and Calabi-Yau geometries.

The arXiv overview gives the standard behind that count (Sections 3.3 and 3.4, Table 3). A closed form is accepted when it agrees with independently computed values to at least thirty digits at kinematic points that entered no stage of the derivation, passed at two working precisions; many are tested to sixty or a hundred digits. For the fifteen known integrals, the published forms were used only at the final comparison, and all fifteen agree. The fifteen new ones have "no prior computation that we could find." The paper is explicit that this evidence is numerical: such integrals "do not constitute a symbolic proof" (Section 3.3). Its derivations are in companion manuscripts posted on bootloops.ai rather than on arXiv.

Projects in other fields

Claude kept matching integrals and algorithms from Schwartz's field to problems elsewhere (section "I know Kung Fu"). Outside his expertise, he writes, he found himself agreeing with Claude's claims, so he brought in experts: "in almost all cases, Claude was technically correct, but the result was not all that interesting until the expert helped steer us." Two projects are described in detail.

  • Ecology. Claude recognized an equation from 2005 by the ecologist Rampal Etienne, which put Stephen Hubbell's neutral theory of biodiversity in testable form but, the post says, nobody could solve at scale for 20 years. Applied to the forest plot on Barro Colorado Island in Panama, the result was that the mix of tree species changes 4.5 times faster than neutral theory allows. The plant biologist James O'Dwyer, an expert in neutral theory, judged that this would likely "be met with a shrug by many ecologists," since the shortfall was already known qualitatively; he proposed subtracting the neutral prediction and modeling the remainder, and the three (Schwartz, O'Dwyer and Claude) built a successor model of forest life histories that the post describes as in excellent agreement with data. The project's web summary words the first result as change about as fast as drift would produce if trees were replaced "four times as often as they are" (two to six times the neutral expectation for most species), and cites a 2014 study (Chisholm et al., Ecology Letters) that had already shown faster-than-drift change at Barro Colorado; the exact paper behind that result is listed as in preparation. The successor model is a bioRxiv preprint by Schwartz and O'Dwyer, A minimal model of a forest after neutral theory (posted 27 October 2026, not peer reviewed). Its author contributions say the two authors conceived the study, chose its questions and tests and wrote the paper; modeling, simulation and analysis were "carried out by M.D.S. with an AI research assistant (Claude Fable 5.1, Anthropic)," which acquired the data, wrote and ran the code and drafted text and figures that the authors revised.
  • Population genetics. Claude evaluated what the post calls a 30-year-old integral expression for how selection shapes rare mutations and applied it to the gnomAD catalog. Schwartz wrote to three biologists before one, his colleague Michael Desai, who works in the field "though not on this exact problem," responded; Desai was impressed by the technique but not the science, and suggested studying pairs of mutations on one chromosome instead. Analyzing 5.7 billion pairs of nearby mutations from the 1000 Genomes Project, they found evidence for gene conversion, a mechanism that the post says nearly every analysis of linked variation ignores. The project's web summary adds the sample (113 Gambian genomes), says a model with gene conversion accounts for about 95 percent of the pairs' departure from the standard picture, and reports that in Zambian fruit flies the same test still fails with conversion added.

The post then lists projects that drifted from exact calculation to broader "Claude-shaped" work:

  • Economics. With two economists, an AI data editor ported the replication packages of 4,452 papers in five journals to open-source code (some 30,000 routines) and checked published numbers against the tables. The NBER working paper (w35782, September 2026, by Schwartz, Isaiah Andrews and Jesse M. Shapiro) reports in its abstract that the workflow flags discrepancies in 3,460 articles or their appendices, cuts computation time by more than a factor of 10 at similar or greater accuracy in 496, and develops an extension aligned with the original article in 923. In the paper (Section 3.1), 35.6% of the calculation discrepancies agree to within one unit of the last printed digit, "consistent with rounding or truncation." Discrepancies were checked by "an adversarial LLM agent" trying to overturn them and, where licensed software was available, by rerunning the original code; the paper reports no human check of a sample. It calls its values current "as of this release" and says "work is ongoing." The NBER copy notes that working papers have not been peer reviewed.
  • Linguistics. With three linguists, Claude compiled AccStack, a database of word stress covering 6,072 languages with the deciding passage quoted for nearly every entry, and a bibliography of 160,000 phonology works. On 13 November 2026 the AccStack site lists 6,080 languages and 160,832 sources and labels itself a beta in which "Everything here is new and still being checked."

A further list of highlights (phylogenetics, the Great Oxidation Event, single-cell gene expression, sunspots, cosmology, mixture models, and Watson's lattice integral) is described as done with experts and "undergoing further exploration and verification." One of these rests on more than the overview reports. The post says "We solved Watson’s “final problem”": the exact return probability of a 3D random walk with three unequal hopping rates. The arXiv overview (Section 5.2) reports only part of that: a proof by computer algebra that no product of complete elliptic integrals represents the generic case, which is a period of a K3 surface. The closed form itself, in genus-two hyperelliptic integrals, is given on the project's web summary; the manuscript (Elkies, Grimm and Schwartz) is listed as in preparation.

Scale, setup and failure modes

Across three months the work produced, in the post's count, 36 manuscripts in 18 fields with 19 coauthors, selected from some 400 candidate problems (section "The technical details"). The manuscripts page dates the projects to July to September 2026 and says all manuscripts "have been read by humans for accuracy," while some drafted by models "have not been fully checked yet by humans" and are watermarked as preliminary; several entries are marked in preparation. Of the 36 entries, the page links one to arXiv (the overview) and others to the NBER paper, the bioRxiv preprint and one further arXiv paper; the rest are PDFs on bootloops.ai.

The setup was one Claude Code session per project, a coordinating session that validated results, and separate sessions for writing, tools and adversarial refereeing; background subagents confined blocks by Fable 5's classifiers to a single agent. The problems Schwartz reports are anecdotes from his own use: no sense of time (after three days' work Claude announced "Two years of campaign, four deep marches," and its estimates were always too long or too short); a preference for grinding through multiday calculations over building a faster tool; context lost to compaction; and declaring victory early, once with a proof complete up to "one unproven lemma" ("That lemma was the whole proof!"). He looks at every plot himself, warns that "the automated checks still can’t be trusted," and supplies the judgment of what is interesting. The projects were compute- and token-intensive; the post gives no figures.

The instructions used

The post and its companion describe the text used to steer the model, without testing it. In the post, Schwartz reports that giving Claude "clear and rigid standards for what success looks like can help" against premature claims of success, and that telling Claude not to commit to a time estimate until it had run a test "didn’t work well either." The arXiv overview (Section 2.3) says the toolkit ships with plain instruction files that a model is meant to read first, stating what each tool does, when to use it and how its answer is tested, along with working protocols, for example to try a method on planted answers before real data. The website's overview page, under its principles, adds that predictions are sealed with a cryptographic hash before results are computed. None of the documents compares outcomes with and without these instructions; their effect is reported as the author's experience.

The author's outlook

Schwartz expects progress through the familiar loop of data and questions, "sped up dramatically," and does not "see any evidence or need to revisit the scientific method." He argues that much technical work can now be automated once the right problem is found, that humans are still needed for the conceptual part, and that the question of credit "won’t be resolved by pretending the human contribution was the typing."

Limitations and open questions

  • Verification is mostly the authors' own. The integral results rest on high-precision numerical agreement at held-out points, a standard the overview states and argues for, but which it says is not a symbolic proof. Most manuscripts are self-hosted, some watermarked preliminary; as of 13 November 2026 the overview is arXiv v1 only, the forest model and the NBER paper are unreviewed preprints, and no journal publication of any project was found. The other-field results were judged by the named collaborators, who are coauthors; in the economics project the discrepancy checks were by a second model and reruns, not by people.
  • Selection. The 36 manuscripts came from about 400 candidate problems; the post does not report how many attempts failed or were dropped.
  • Novelty claims are the author's. "Never before computed" rests on the authors' literature search ("no prior computation that we could find"). In ecology the post itself reports that the headline finding was already known qualitatively.
  • Differences between the post and companions. The post's 4.5-fold figure appears in the web summary as replacement "four times as often"; the 4.5 itself was not found in any document read. The Watson closed form is on a web page, with its manuscript in preparation. AccStack counts have changed since publication.
  • Interest. Schwartz was a visiting researcher at Anthropic, and the arXiv overview, the forest preprint and the website state that Anthropic funded the project; the post appears on Anthropic's blog and concerns its model. A search found news coverage of BootLoops but no published critique of its results; none was read for this note.
  • One supervisor, one model generation. The account is one expert's experience with Claude Fable 5 (the overview credits Fable 5.1 with assisting its writing), with no comparison to other models or supervisors.

Research questions

  1. How many of the fifteen new integrals are independently confirmed? The acceptance standard is numerical agreement to thirty or more digits. Independent recalculations by other groups, or symbolic proofs, would show whether the standard misses errors in practice.
  2. What fraction of the 400 candidate problems produced a manuscript, and why did the rest stop? A log of attempts by outcome (failed, uninteresting, dropped for time) would show the selection behind the 36.
  3. How often are the NBER workflow's flagged discrepancies real errors? The paper counts 3,460 articles with flagged discrepancies and puts 35.6% of calculation discrepancies within rounding of the last printed digit, with checks by a second model. A hand-checked sample of the rest, classified as errors in the original, porting artifacts or conventions, would show what the count measures.
  4. How much did expert steering change the result in each project? Comparing Claude's first finding with the final manuscript, project by project, would measure the experts' contribution that the post describes qualitatively.
  5. Do the harness's instruction files and protocols change outcomes? Running the same tasks with and without the instruction files, planted-answer protocol or success standards, and counting false claims of completion, would test what the post reports from experience.
  6. Do the cross-field results hold up in their own fields? Peer review or replication of the gene-conversion, forest and phylogenetics manuscripts by specialists who were not coauthors would answer this.

How it connects

References