version: "1.0.0" name: test-driven-research-code description: Apply lightweight test-driven development to research code so analysis logic is trustworthy. Write a failing test for each transform or metric before implementing, and guard known-correct values. Use when writing data-cleaning, feature, or metric code whose correctness affects a reported number. Trigger on "test this function", "is the metric right", "regression test", "guard the value". Pairs with reproducibility and statistics. Structure adapted from agent-skills (MIT). See /ATTRIBUTION.md.

Test-driven research code

When to use

Data transforms, metric and estimator implementations, anything feeding a headline number.

Core process

1. Write a failing test with a hand-checked expected value before coding the transform.
2. Implement minimally until it passes; keep the test as a regression guard.
3. Add an edge-case test (empty, NaN, single row, off-by-one window).
4. Run the suite in the make or CI path so it gates every rerun.

Anti-rationalization table

Exit criteria

• [ ] Each result-feeding function has at least one value-checked test, run in CI.
• [ ] The statistical correctness of the chosen metric is confirmed with the statistics council.

Red flags

• A metric implementation with no test. Do not let its output become a reported number.
• A test that asserts the code's own output rather than an independently known value.