BootLoops-ai/bootloops

What it solves

LLMs are often used for quantitative science but their outputs are hard to trust without verification. BootLoops solves this by providing a toolkit of scientific instruments plus strict protocols that make results checkable. Each tool is documented for an agent: what it does, when to use it, what its output means, and what test the answer must pass.

How it works

You clone the repository, start your agent inside it, and ask it to read the index (tools/README.md) and propose a plan. The agent selects relevant instruments from the 50 packages under tools/; each package has a GUIDE.md with purpose, acceptance gates, and verification class. The toolkit includes tools for multidimensional integrals, recurrences with certificates, Bayesian evidence, ball arithmetic, enumeration with completeness proofs, and more. External engines (Kira, FireFly, FORM, etc.) can be installed on demand; three engines (SOFIA.jl, Eichler.jl, Leviathan) ship in full source. A one-command selftest (python3 run_selftests.py --par 8) validates a fresh clone in minutes.

Who it’s for

Researchers and engineers who want to leverage LLM agents for reliable quantitative computation—especially in mathematical physics, statistics, and related fields. It is also for anyone who wants to contribute tools or protocols to grow the harness.

Highlights

• Model-agnostic: works with any agent framework (Claude, Codex, etc.)
• 50 self-contained Python packages with per-package batteries (selftests)
• Includes Julia ports and patched forks of major engines (Kira, Blade, AMFlow.cpp)
• Skills repository (plain-markdown protocols) defining rigorous checks: reproducibility, positive controls, integer-relation discipline
• Entire toolkit runs on a laptop with minimal dependencies (mpmath, sympy, numpy, python-flint)
• Clear licensing and attribution, with third-party notices