Matthew Schwartz credits 15 new physics integrals to a harness that keeps Claude on bounded, checkable tasks
Harvard physicist Matthew Schwartz says his BootLoops toolkit helped Claude finish 30 scattering-amplitude integrals, 15 of them never completed before. His examples show the harness keeps the math checkable while collaborators still decide which questions are worth computing.
The Engineer · Build desk
What happened
- Anthropic ran the guest post on October 1, and BootLoops itself is Schwartz's own open-source project, with its code on GitHub.
- Schwartz calls the core problem an "impedance mismatch": scientists approach models as collaborators, while current systems do best at bounded work such as coding, parsing papers and structured calculation.
- In population genetics, the team analyzed 5.7 billion pairs of nearby mutations from the 1000 Genomes Project and reports evidence of gene conversion.
- Schwartz describes the ecology, genetics and other cross-field projects as ongoing work with collaborators, with more still under exploration and verification.
Why it matters
- What it limits constraint A team that funds only the checking layer gets verified answers to questions the model chose; Schwartz found those were often scientifically unremarkable until a specialist reshaped them.
- Who has to decide decision BootLoops is open source and built for different models, so a lab can trial the harness without switching vendors; specialist time to pick questions has to be budgeted separately.
- Who's exposed exposure Anyone justifying adoption on the 15 new integrals is citing Schwartz's own count from a post on the model vendor's blog, with no independent evaluation behind it yet.
Next story · Invest · Oct 9
OpenAI and Anthropic executives war-game the backlash from an AI cyberattack insiders expect within a yearConfirmed4 publishers
Schwartz's answer to the mismatch is to pick what the post calls "Claude-shaped" problems and build a harness around the model [5]. In his account, the harness is a set of computational tools Claude can use to port, codify and verify mathematical physics calculations [8]. He assembled it from software and protocols written while using Claude on his own research [9]. To run it, he coordinated several projects through separate Claude Code sessions, with intermediate work kept in files [18]. Neither write-up describes the tool interfaces or what a given check tests.
The headline count splits evenly. Of the 30 scattering-amplitude integrals Schwartz reports, 15 reproduced known results and 15 had not been completed before [4]. So half the output was the pipeline proving itself against answers already known [20]. I think that ordering is the part to copy. Establish the loop on known answers, then point it at open ones. The count is Schwartz's own, given in Anthropic's post, and runtimewire notes it is not an independent evaluation of the toolkit [6]. For the result to carry over to another group, that group needs a stock of known answers to reproduce and results exact enough to check mechanically. BootLoops was built for exact computation [1].
The cross-field work shows where checking stops helping. Schwartz writes that Claude's suggestions there were often technically sound but scientifically unremarkable until researchers with the relevant expertise shaped the question [10]. In the forest project with plant biologist James O'Dwyer, the first result showed tree-species composition at Panama's Barro Colorado Island changing faster than a neutral-theory model predicted [11]. O'Dwyer judged that finding alone would draw little interest from ecologists. He steered the work toward a model that better characterizes species' life histories [11]. Schwartz says that model matched data and is being extended to other forest plots [12]. In genetics, collaborator Michael Desai moved attention from an initial calculation toward relationships between mutation pairs [13].
Inside the loop, Claude still failed in ways Schwartz lists. It declared tasks finished before resolving the central problem, misjudged how long work would take, and pursued long calculations where, by his account, building a new tool was the better route [15]. Anyone who has reviewed a junior engineer's "done" ticket will recognise the first. In my view a harness handles that one directly, provided "done" means an exact result that passes a check. Building the new tool stays with the researcher in Schwartz's process, where he develops the tools and representations that make the work tractable for the model [7].
Agents did the calculating in those Claude Code sessions [18]. The record therefore supports building checks around agents. Runtimewire puts the human share as choosing what to pursue and judging whether an answer mattered, beyond checking the model's math [14].
What to watch
- Independent replication of any of the 15 previously uncompleted scattering-amplitude integrals, which would turn Schwartz's count into a checked result.
- Publication of the Barro Colorado forest model and the gene-conversion analysis, both described as ongoing work.
- Reports from groups running BootLoops on non-Claude models, testing whether the harness works across models as designed.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence40
- Adoption10
- Hype gap+15
- Incentives60
- Confidence55
Perspective Coverage
3 publishers- Builder
- Builder 40%
- Operator
- Operator 42%
- Investor
- Investor 18%
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
Anthropic published a Science Blog guest post by Harvard physicist Matthew Schwartz introducing BootLoops, a toolkit for exact computations in quantitative science, as a way to narrow the gap through guided, verifiable workflows rather than end-to-end automation.
- [2]
Anthropic's October 1st post describes Schwartz's BootLoops, an open-source toolkit; BootLoops is Schwartz's project, not an Anthropic product, and is available on GitHub.
- [3]
Schwartz calls the problem an "impedance mismatch": scientists often approach models as human collaborators, while current systems are better at bounded tasks such as coding, parsing papers and performing structured calculations.
- [4]
Schwartz reports 30 scattering-amplitude integrals completed using the workflow: 15 reproductions of known results and 15 calculations he says had not previously been completed.
- [5]
The post focuses on finding "Claude-shaped" problems and creating a harness around the model, the harness being BootLoops.
- [6]
The integral figures describe Schwartz's account in Anthropic's guest post; they are not, by themselves, an independent evaluation of the toolkit's performance.
- [7]
Schwartz's account describes an iterative process in which Claude is steered as a research assistant while the researcher develops tools and representations that make the work more tractable for the model.
- [8]
Schwartz describes building computational tools the model can use to help port, codify and verify mathematical physics calculations.
- [9]
Schwartz says he assembled the toolkit from software and protocols developed while using Claude on research calculations.
- [10]
Schwartz writes that Claude's cross-field suggestions were often technically sound but scientifically unremarkable until researchers with relevant expertise helped shape the question.
- [11]
With plant biologist James O'Dwyer, the first result showed tree-species composition at Panama's Barro Colorado Island changed faster than a neutral-theory model predicted; O'Dwyer thought that finding alone would draw little interest from ecologists and helped redirect the work toward a model that could better characterize species' life histories.
- [12]
Schwartz says the resulting forest model matched data and is being extended to other forest plots.
- [13]
In genetics, collaborator Michael Desai helped direct attention from an initial calculation toward relationships between mutation pairs.
- [14]
The human contribution was choosing what to pursue and assessing whether an answer mattered, not simply checking the model's arithmetic.
- [15]
Schwartz describes failures: Claude could declare a task finished before resolving a central problem, misjudge how long work would take, and pursue long calculations where building a new tool would have been the better approach.
- [16]
In population genetics, the researchers analyzed 5.7 billion pairs of nearby mutations from the 1000 Genomes Project and found evidence of gene conversion, a process Schwartz says is commonly left out of analyses using linked genetic variation.
- [17]
The post describes the genetics and several other projects as ongoing work with collaborators; Schwartz says additional projects are undergoing further exploration and verification.
- [18]
Anthropic's post says Schwartz coordinated multiple projects through separate Claude Code sessions, with agents running calculations and intermediate work organized in files.
- [19]
BootLoops is designed to be used with different models, according to Schwartz's post.
- [20]
Half of the 30 integrals Schwartz reports were reproductions of known results.
Sources
3 independent publishers whose own reporting we read for this story.
- dev.toBootLoops Shows How AI Can Bridge the Science Impedance Mismatch
1 article · October 1, 2026
- runtimewire.comAnthropic publishes Matthew Schwartz's toolkit for AI-assisted science
1 article · October 1, 2026
- the-decoder.comOpen-source "BootLoops" harness supports AI models in performing precise scientific calculations
1 article · October 3, 2026
Topics and entities
Topics
- AI-assisted science
- Agent Harnesses
- Computational physics
Entities
- 1000 Genomes Project
- Claude
- Harvard University
- Matthew Schwartz
- Claude Code
- Anthropic