BootLoops-ai/bootloops

BootLoops

BootLoops 1.0 is a harness for large language models doing precision quantitative science. At its core it is an extensive scientific software package (written, ported, and upgraded to be driven by an LLM) together with the working protocols that make the model's results checkable. The harness is independent of the model driving it: clone it, point whatever agent you use at it, and the agent gains instruments it can run. Each is documented in the terms an agent needs: what the instrument does, when to reach for it, what its output means, and what test its answer must pass before anyone believes it.

The kinds of things it does:

• Frontier integrals of mathematical physics — multi-dimensional

parametric integrals whose answers live in the polylogarithmic, elliptic, K3, and Calabi–Yau classes, in closed form or to hundreds of certified digits.

• Recurrences with proofs — the equation a sum or integral provably

satisfies, certificate included: strong enough to prove a closed form impossible, not merely fail to find one.

• Bayesian evidence integrals in closed form — the marginal likelihoods

that decide model comparison, integrated analytically or with certified quadrature instead of sampled.

• Monte Carlo replaced by convergent methods — quadrature and recurrences

that carry certified error bounds instead of statistical ones where the problem admits them; where it does not, the sampling tools state statistical error bars as such.

• Ball arithmetic for effective error propagation — every value carries

a proven error radius through every step, so rounding and truncation error can never go unreported.

• Exhaustive enumeration with completeness certificates — check every

case and prove that none was missed.

• Open implementations of standard statistical procedures — routines

specified by public documentation and the underlying methods papers, written from scratch in open code and validated against published outputs.

• Comprehensive field-specific codebases — JaCKandJill (Bayesian

phylogenetics, in its own repository), Mixalot (statistics), Terrier (the string landscape), Popcorn (population genetics).

One core runs through all of these: exact reduction, singularity analysis, certified transport, evaluation to hundreds of digits — machinery that applies wherever a computation ends in a number someone needs to trust. The toolkit was built computing scattering amplitudes; that story, and the results, are told at bootloops.ai. The package itself is general purpose.

The harness is meant to grow. Every problem it meets leaves tools behind for the next one, and additions are welcome from anyone: submit a pull request, suggest a tool addition or upgrade, or point us at independent code you want ported or linked — through this repository's issues or the contact page at bootloops.ai.

The skills are the other half: the protocols, as plain-markdown instruction files any agent framework reads, kept in their own repository, skills (see Related repositories below). They encode what "done" means: a result reproduces an independent route at points no fit ever saw, with a positive control proving the check can fail; integer-relation discipline with the constant ring declared before the search and refusal over invention when no relation is found; planted-truth controls that recover a known answer before any real data is touched; provenance bookkeeping so no oracle that fed a fit ever certifies the result; timing discipline (measure a small run before a big one); and the heavier protocols for proving with agents, simulating referees, auditing the literature behind a novelty claim, verifying bibliographies, and editing prose for precision and honesty (no unsupported claims, no filler, every quoted number traceable).

Using it

Clone this repository, start your agent inside it, and put your problem to the model with the toolkit in front of it:

Read tools/README.md and the tool guides it points to, then propose a plan

for this problem for me to approve.

Name the object if you can: an integral from a paper, a dataset, a published number you want checked. The model reads the index, says which instruments apply (or that none do), and you approve the plan before anything runs. The tools are cheap to drive: a laptop and whatever model access you already have are enough.

Validating the install

One command verifies a fresh clone:

python3 run_selftests.py --par 8

This walks every package under tools/ (49 of them), runs each package's own battery on the toy data it ships with, and prints a status line per package. The run takes a few minutes on a laptop: the packages come back green, and 3 stop with a named error because they need data the repository does not include. If a battery needs an external engine you have not installed, it skips and says what to install. For a first real computation, the one-loop box through the Landau Alphabet engine:

python3 tools/landau-alphabet/test_landau_alphabet.py

reproduces the solver's reference results in about a minute and a half.

Repository map

• tools/ — the toolkit packages, one directory per package,

each with a GUIDE.md or README.md (purpose, acceptance gates, verification class). The per-package index is tools/README.md.

• toolkit/ — the rosters and recipes:

ours/README.md (the tool roster by theme), ours/RECIPES.md (working recipes with their checks), external/TOOLS.md (the external engines, with licenses and where to obtain each).

• upgrades/ — the house engines, shipped in full

source: SOFIA.jl (a Julia translation of the SOFIA package by Correia, Giroux and Mizera, with no Wolfram dependency and their original files under reference/), Eichler.jl, and the Leviathan Landau engine. Each directory carries PATCHES.md and its license. The patched forks of Kira, Blade, and AMFlow.cpp live in their own repositories (below).

• ops/ — operations packages: infrastructure for running

the toolkit on a shared machine rather than scientific instruments. Currently Turnstile (ops/turnstile), admission control for long jobs — a priority token plus RAM and CPU-width ledgers — with its own self-test.

Related repositories

BootLoops is six repositories published side by side at github.com/BootLoops-ai. This one, bootloops, holds the toolkit and the house engines; the other five are listed below. Where the code or the guides assume a location for a sibling repository, it is a checkout beside this one (../ relative to this repository's root).

• jackandjill —

JaCKandJill, exact and certified Bayesian evidence for phylogenetic models (Python package phyloexact, with its own self-certification suite). MIT.

• amflow-cpp — the

patched fork of AMFlow.cpp (precision fixes, Kira/FireFly robustness, checkpoint/resume) with the validation wrapper suite in bootloops-wrappers/; PATCHES.md records every change against the AMFlow.cpp authors' code. MIT.

• kira — the patched fork of

Kira 3.1 (parallel back-substitution, SQLite autocommit guards, the 128-bit-weight build variant); PATCHES.md and PATCHES.diff record the delta, and bootloops-tools/ holds two GPL Python utilities derived from Kira's and FireFly's file layouts that kira-stack uses optionally. GPL-3.0-or-later.

• blade — the patched fork

of Blade rebuilt to run without a Wolfram kernel, the Blade authors' MIT license intact (plain-C FiniteFlow wrappers, a CLI, a sample-point dumper); PATCHES.md there. The Python port of Blade's search logic that drives those binaries is in this repository, under tools/blade/.

• skills — the working

protocols as plain-markdown Agent Skills readable by Claude Code, Codex, Cursor, Copilot, and most other agent frameworks, with their packaging as installable Claude Code and Codex plugins.

Requirements and tested platforms

The toolkit is plain Python 3.12 (a few packages carry Julia components) with a small dependency core: mpmath, sympy, numpy, python-flint, and pytest for the batteries. Julia packages ship manifests generated on Julia 1.11; Julia 1.12 works after a Pkg.resolve(). The release was tested on:

• Linux x86_64: Ubuntu 24.04 class (the development platform).
• Linux containers, x86_64 and arm64: Debian 12 with Python 3.12 and

Julia 1.11: the full validation run passes in a fresh minimal container on both architectures. One known arm64 limit: the Blade fork's from-source installer (blade) stops inside MPFR's own test suite on arm64; the full Blade build is verified on x86_64.

macOS is not part of the release testing. The pure-Python packages have no platform-specific code and the Blade fork ships a macOS installer, but run the validation command above before relying on anything there.

External engines

Several packages drive external programs (Kira, FireFly, FORM, AMFlow, FLINT, msolve, Singular, OSCAR among them). The house engines ship in this repository in full source under upgrades/; the patched forks of Kira, Blade, and AMFlow.cpp build from their own repositories (see Related repositories above and INSTALL.md); the rest are not downloaded automatically: you (or your agent) install them when a tool needs one. Nothing breaks in the meantime: a battery whose engine is absent skips with a message naming exactly what to install, and toolkit/external/TOOLS.md lists every engine with its license and where to obtain it. A practical route is to let the model handle it: *"the selftest says FORM is missing; install what it asks for."*

Installing

See INSTALL.md: clone the repository for the toolkit and the house engines; the skills install from the sibling repository skills by clone or through the plugin routes described there. The toolkit packages are plain Python (plus Julia for a few); path setup and the external-engine prerequisites are in the same file.

The papers, the result pages, and the per-problem code live at bootloops.ai; this repository is the harness itself.

Status, provenance and responsible use

The code was written by Claude working under the author's direction: the author set every problem, approved every plan, and checked the reported results against independent routes; each package ships the acceptance battery described in its GUIDE so users can re-verify.

These are research instruments. Validate outputs before relying on them; nothing here is intended or fit for clinical, actuarial, payment, regulatory or public-safety decisions.

Acknowledgments

BootLoops is a harness around other people's mathematics and software, and it is a pleasure to say whose.

Exact arithmetic and computer algebra. Every certified digit here passes through FLINT and Arb (William Hart, Fredrik Johansson, Albin Ahlbäck and the FLINT developers; acb_theta by Jean Kieffer) via python-flint (Fredrik Johansson, Oscar Benjamin), Nemo/Hecke and Arblib.jl, alongside mpmath, SymPy, SageMath, ore_algebra, Singular, msolve, OSCAR, PARI/GP, GiNaC, fplll and Julia.

Loop-integral engines. Kira (Philipp Maierhöfer, Johann Usovitsch, Peter Uwer, Jonas Klappert, Fabian Lange, Zihao Wu) with FireFly (Jonas Klappert, Sven Yannick Klein, Fabian Lange) and Robert H. Lewis's Fermat; the auxiliary-mass-flow method and AMFlow (Xiao Liu, Yan-Qing Ma and collaborators) and AMFlow.cpp (its contributors, maintainer @chang18); Blade (Xin Guan, Xiao Liu, Yan-Qing Ma, Wen-Hao Wu) on Tiziano Peraro's FiniteFlow; FORM (Jos Vermaseren and the FORM developers), HyperFORM (Adam Kardos, Sven-Olaf Moch, Oliver Schnetz) and Erik Panzer's HyperInt; pySecDec (Sophia Borowka, Gudrun Heinrich, Stephen Jones et al.); FIRE and FIESTA (Alexander Smirnov et al.); SOFIA, PLD and SubTropica (Miguel Correia, Claudia Fevola, Mathieu Giroux, Sebastian Mizera, Giulio Salvatori, Simon Telen) with Effortless (Antonela Matijašić, Julian Miczajka); Landau's Leviathans and SPQR (Vsevolod Chestnov, Giulio Crisanti, Mathieu Giroux); PentagonFunctions (Dmitry Chicherin, Vasily Sotnikov, Simone Zoia). The Landau-bootstrap and sequential-discontinuity tools grew out of work with Holmfridur Hannesdottir, Andrew McLeod, Cristian Vergu and Jacob Bourjaily; the lattice-reduction regression out of work with Oscar Barrera, Aurélien Dersy, Rabia Husain and Xiaoyuan Zhang.

Applied packages and data. The population-genetics, ecology, seismology, phylogenetics and public-records tools stand on dadi, fitdadi, polyDFE, fastDFE, moments, msprime/tskit, etas, matPTF, SageMath's genus-2 modules and the papers named in each GUIDE.md; the datasets of Dmitry Chicherin, Ian Moult, Emery Sokatchev, Kai Yan and Yunyue Zhu and of the DravLex team (Vishnupriya Kolipakam and co-authors) are used under CC BY 4.0 with thanks.

Full author lists and references are in toolkit/external/TOOLS.md, REFERENCES.md and each package's CREDIT paragraph; please cite those authors, not only BootLoops. Omissions are ours to fix: open an issue and they will be.

Maintenance, reporting and security

Maintenance. This repository is maintained by Matthew D. Schwartz, not by Anthropic. It is not an officially supported Anthropic product, and Anthropic does not provide support, updates or fixes for it.

Reporting issues. Please report bugs and security problems through this repository's GitHub issues.

Security considerations. Treat input files from others as code. These are research tools meant to be run locally on inputs you trust. Many of them evaluate the contents of their input files (JSON, YAML, .m, .ms, .jl, pickle and similar), so a file received from someone else can run arbitrary commands on your machine. Only run files you wrote yourself or got from a source you trust, or run them in a sandbox or container. The integrity checks and certificates in this repository guard against accidents. They are not a security boundary.

License and attribution

BootLoops 1.0 is released under the MIT License (LICENSE), Copyright (c) 2026 Anthropic, PBC. Created by Matthew D. Schwartz; code written by Claude (Anthropic) under his supervision. This is not an officially supported Anthropic product; it is maintained by Matthew D. Schwartz (https://www.bootloops.ai). The prose and figures written for this repository are released under CC BY 4.0 (LICENSE-CONTENT). The reference copies of third-party sources under upgrades/SOFIA.jl/reference/ and tools/subtropica/reference/, and the Python port of Blade's search logic in tools/blade/ and the Julia port of SubTropica's front end in tools/subtropica/, keep their original licenses and copyright notices (MIT under their authors' copyright); three files are GPL: tools/subtropica/src/lr_refine.jl (with its test), derived from Erik Panzer's HyperInt, GPL-3.0-or-later (see tools/subtropica/NOTICE); tools/eichler/genus2/mestre_port.py, a sympy translation of SageMath's mestre.py and invariants.py by Florian Bouyer, Marco Streng and Nick Alexander, GPL-2.0-or-later (see tools/eichler/NOTICE); and tools/formglue/form_hyper.py, whose HyperFORM driver template follows the example drivers shipped with HyperFORM by Adam Kardos, Sven-Olaf Moch and Oliver Schnetz, GPL-3.0-only (see tools/formglue/NOTICE). The patched engine forks are distributed in their own repositories under the same organization, each under the license of its original authors (Kira GPL-3.0-or-later; Blade and AMFlow.cpp MIT), and are not part of this repository. THIRD_PARTY.md has the full list and NOTICE the attribution.

To cite BootLoops: M. D. Schwartz, BootLoops 1.0 (2026), bootloops.ai.