AI-assisted development

Parts of this codebase, its tests, benchmark drivers and documentation were written or revised with AI assistance: Anthropic Claude models through Claude Code (including agentic sessions that implemented driver extensions, ran the benchmark drivers on the maintainers’ compute pool and drafted the resulting documentation), and OpenAI GPT-5-series models through OpenCode. Pull requests are reviewed by the maintainer and by automated code-review services (OpenAI Codex and CodeRabbit). Every change is gated by the test suite and CI before it reaches main (a pull request stacked on a parent branch gets that run when its stack lands), and benchmark results carry metadata sidecars that pin the generating commit and environment; the maintainers review and remain responsible for all code and claims in this repository.

---

The mechanics behind that paragraph are documented in the rest of this section: the gates are CI and the review bots, what a measurement has to record is Benchmarks and reproducibility, and what a contribution has to carry — whoever or whatever drafted it — is Contributing.

One scope note, since the statement above is deliberately kept as written. The metadata guarantee is about promoted results — a full run wrapped by tooling that pins the generating commit and the environment — not about every run that happens to leave a file behind. The drivers those results came from are no longer in this repository; the record a measurement still has to carry is Benchmarks and reproducibility.