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