These requirements apply to the current ML4T libraries and the maintained Zipline, Pyfolio,
Alphalens, and Empyrical Reloaded repositories. Use the repository that owns the affected code.
Shared policy and qualification belong in ml4t/ecosystem; library changes belong in that library.
Search existing issues before opening a new one. Each issue describes one problem or coherent change. Open separate issues for unrelated problems.
For a bug, include the affected package version, environment, expected behavior, observed behavior, and the smallest complete reproduction. Report suspected vulnerabilities through the private process in SECURITY.md.
Discuss substantial changes in the owning issue before implementing them. Reloaded libraries accept bug, security, and compatibility fixes; direct new feature requests to the current ML4T libraries.
For external contributions, each pull request normally addresses one issue and includes a closing
reference, such as Closes #123. Maintainer-directed batches follow the exception below.
Related code, tests, and documentation can belong in the same PR. Keep unrelated fixes, refactors,
and formatting changes separate. Automated dependency updates and administrative metadata changes
do not require a separate issue.
For a bug fix, include a regression that exercises the affected public API or user workflow. Run it against the relevant base revision and the proposed fix, and record both revisions, commands, and results. It must fail because of the defect on the base and pass with the fix. Tests of copied logic or source text do not establish that the behavior is corrected.
Run the owning repository's documented quality checks before requesting review and report the commands and results. Documentation changes need documentation checks; they do not need an artificial regression test. This repository's commands are in AGENTS.md.
Verified defects must be corrected, including defects that affect numerical results. If a proposed change alters an intended definition, annualization, sign convention, or default, agree on the expected behavior in the owning issue and document its compatibility impact. Existing output alone is not evidence that the implementation is correct.
Our own fixes, including work delegated to agents, should use bounded, coherent PRs. A PR may close several related issues when they share a user workflow, subsystem repair or compatibility objective and can be reviewed and reverted together. Include implementation, regressions, docs and follow-up repairs in that PR. Keep individual defects and their evidence traceable through issue references and focused commits; do not create an issue or PR for each implementation step.
Split unrelated changes, independent release or rollback decisions, separate ownership boundaries, and independently risky or breaking changes. Avoid both tiny fragments and batches whose scope cannot be reviewed clearly. A general maintenance label is not sufficient justification to combine work. Reuse existing submissions where coherent; consolidate owned branches only while preserving contributor attribution, recoverable history and other owners' active work.
Run focused checks during iteration and batch related work before expensive CI. The complete final revision still needs substantive review and every applicable required check. Batching does not waive failing-base/passing-fix evidence, AI disclosure, compatibility review or current-revision qualification.
You may use AI tools for code and prose. State whether you used them and which parts they helped produce. Review and test the submission yourself, and be able to explain the code, evidence, and responses you submit. An automated reply does not replace understanding a review question.
Keep concurrent submissions manageable. Wait for feedback on your first focused contribution before opening a large series of similar PRs.
Maintainers may ask you to split or complete a submission, or close it without detailed review when it does not meet these requirements. You are welcome to resubmit once it does. Existing bundled PRs may contain useful fixes; maintainers can request focused replacements so each fix can be assessed independently.