🤖 LLM Transparency

This page describes how LLMs were (and continue to be) used on this project, so that readers can weigh the codebase, the documentation, and the accompanying paper with the right context in mind.

We highly respect that readers expect human-written content. The majority of texts in this repository are human-typed; LLMs are used as authoring and coding aids under iterative human direction, not as fully autonomous authors.

Codebase

A large portion of this codebase was written with GitHub Copilot in the early stages. Nearly all subsequent coding has been carried out through vibe coding with Claude Code and Codex since they became available. All has been human-reviewed by the author before being made public.

The author does understand the algorithmic and theoretical backgrounds behind the solver and is responsible for the design decisions that shape the codebase, even where the keystrokes were produced by a coding agent.

That said, UI logic and elementary math (vector arithmetic, index bookkeeping, glue code) are not reviewed with the same depth. LLMs are reliable enough here that the author's careful review is concentrated where it matters: the solver's algorithms and the design decisions behind them.

README

Up through March 2026, this README was mostly hand-typed in the author's voice and then proofread by an LLM with minimal changes. Minor parts (e.g., tables) were greatly assisted by an LLM.

Since April 2026, the README and other articles in this repository are written directly by an LLM under the author's instructions, without a prior hand-typed draft. The author still carefully reviews every passage to ensure the wording stays faithful to the author's voice and intent. Corrections are applied wherever the LLM drifts from how the author would have phrased it. The shift is in how the text is typed, not in who is responsible for what it says.

An LLM is also used to proofread and polish the wording, though this iterative process has occasionally introduced minor expansions or contractions; all such changes are carefully human-checked and corrected where necessary.

Python Docstrings and Example Notebooks

Python docstrings are auto-generated by an LLM. Comments in the example Jupyter notebooks are also auto-generated with an LLM.

Both are intended as convenience layers on top of code that the author has written and reviewed; the source of truth remains the code itself.

Paper Draft

The paper draft is written directly in English, not in Japanese and then translated. Writing directly in English avoids polluting the text with hallucinated nuances that translation can introduce. The author has verified not only the surface-level meaning but also the delicate nuances throughout.