# LLM assistance Large language models (Anthropic's Claude, through Claude Code) write much of this code, its tests and its documentation, and we do not hide it. The repository is built so that this is safe to rely on. - **A person reviews every change.** A model proposes a commit; it does not decide one. - **Checks gate the code.** `pixi run check` runs the formatter, the linter with the full rule set, the type checker, the Markdown lint, the pytest suite and the repository's own checker, `paperspec check`. CI runs the same, with branch coverage that fails under 90%, and the stress test. - **The oracle checks itself.** Every expected value in the vectors and fixtures is computed by the code and rebuilt byte for byte, so a change that alters a computed value shows up as a changed file in the diff. - **Assisted commits are marked.** Each carries a `Co-Authored-By` trailer naming the model, and a `Claude-Session` trailer linking the session. - **Claude's working material stays out of the repository.** Plans, notes and review reports live in an untracked `work/` folder. The specification is the authority. Where a model's code and a chapter disagree, the chapter wins and the code is the bug.