# Contributing and independent review

Start with `docs/EXPERT_REVIEW.md`, `AGENTS.md`, and the evidence ledger. The owner
has not yet selected a reuse license; this guide does not grant reuse or contribution
rights or change that decision. Questions and review findings can be shared through
the Hugging Face repository's Community discussions when the owner enables them.

For a reproducibility report, include the archive hash or Hub commit, environment,
command, seed, expected/actual result, and modified files. For a mathematical issue,
include the relevant theorem assumptions and a minimal counterexample. For overlap,
include a precise primary reference and a mapping to the mechanism.

Use `docs/REVIEW_REPORT_TEMPLATE.md`. A useful negative result is welcome: shared-tool
ties, hidden replay, excessive metadata cost, and failed nonlinear extensions each
help locate the architecture's actual limits. Evidence should remain separate from
subjective novelty or importance ratings.

Proposed implementation edits should preserve exact abstention and error semantics,
provide relevant tests, and document changed assumptions. Experiments belong in
`local_runs` until explicitly accepted as new release evidence. No automated command
contacts external experts, creates discussions, or publishes findings.