Hardening Research Code

Hardening Research Code

Make research and scientific code trustworthy: correct, regression-safe, and numerically stable by establishing explicit criteria and backing them with targeted tests.

Interaction mode

This skill leans Direct by default. For the full Collaborative-vs-Direct protocol and override rules, see the Interaction Modes reference in the ai-research-workflows:using-research-workflows skill.

Purpose

Research code often starts as exploratory scripts. Hardening turns it into trustworthy software: outputs are verifiably correct, regressions are caught automatically, and numerical behavior is documented and bounded.

Iron Law: a regression baseline is not a correctness check

Pinning the current outputs proves only that behavior hasn't changed — not that it is right. Never label outputs as a verified or "golden" baseline until they are checked against an independent reference (analytic limit, reference implementation, published result, or invariant).

Before falling back to regression-pinning, you MUST confirm no analytic limit, reference implementation, or invariant is reasonably derivable. If one is, verify against it first — "the output looks plausible" is not verification. If you pin outputs that have NOT been independently verified, label the baseline UNVERIFIED so it is never mistaken for established-correct values, and say so to whoever relies on it.

Establish correctness criteria

Before writing any test, decide what "correct" means for this code:

• Analytical / closed-form solutions — does the code reproduce a result that can be derived by hand or from first principles? (e.g., the discrete Fourier transform of a pure sine wave has a single nonzero bin.)
• Reference implementations — does it match a trusted, established library or codebase on shared inputs?
• Published results — does it reproduce a table, figure, or metric from a paper within stated tolerances?
• Conservation laws / invariants — does a physical or mathematical invariant hold? (e.g., total energy is conserved, probability sums to 1, a rotation matrix has determinant 1.)

Document the chosen criterion explicitly before writing tests. A test without a stated criterion is uninterpretable when it fails.

Techniques

Golden / reference tests

Compute expected outputs from a trusted source — an analytic formula, a reference implementation, or a published dataset — and assert that the code under test matches those outputs within tolerance. Store expected values and any input data in a versioned file alongside the tests.

Numerical-tolerance comparisons

Never use exact float equality. Always use absolute or relative tolerances appropriate to the domain:

# strategy-level pseudocode — defer mechanics to python-testing skill
assert |result - expected| < atol + rtol * |expected|

Document the tolerance choice: what physical or numerical argument justifies it? If tolerance is tightened or loosened later, the justification must be updated.

Choosing the bound: start from the floating-point floor — roughly machine epsilon times the number of accumulating operations (atol ≈ N · eps · scale) — and widen only for a documented reason (iterative-solver tolerance, stochastic variance). Use rtol for values spanning orders of magnitude, atol for values near zero. Never tune the tolerance to make a failing test pass — a bound chosen to fit the result is a silent correctness hole; derive it from the numerics, not the output.

Regression tests pinned to accepted outputs

Use this only after confirming no analytic limit, reference implementation, or invariant is reasonably derivable (see the Iron Law) — a regression baseline is a fallback, not a substitute for verification. When it genuinely is the only option: run the code on canonical inputs, check the outputs against whatever partial criteria you can (ranges, signs, invariants, expected trends), and pin them. Future runs must stay within tolerance of that baseline. If the pinned values were not verified against an independent reference, label the baseline UNVERIFIED so it is never mistaken for established-correct output. Update the baseline deliberately, with a commit message explaining why the output changed.

Stability checks

Perturb inputs, random seeds, data types (e.g., float32 vs float64), or batch sizes, and verify that outputs vary by a bounded amount. A result that changes dramatically under small perturbations is not trustworthy, regardless of how close it is to a reference on the nominal input.

Property / invariant checks

Assert structural properties that must hold regardless of the specific input values: symmetry, monotonicity, conservation, idempotence, or range bounds. These catch whole classes of bugs that value-comparison tests miss.

Deferral

Defer pytest fixtures, parametrization, numpy.testing / torch.testing assertion mechanics, and CI configuration to the scientific-python-development:python-testing skill. This skill is the research-specific strategy layer: what to validate and why. Apply the same principles to R, Julia, Fortran, or any other runtime.

Workflow

1. Pick criteria — choose from the options in "Establish correctness criteria" above; document the choice in a comment or README near the tests.
2. Add the smallest tests that capture them — one test per criterion; avoid testing implementation details.
3. Run — and confirm each test can fail. Confirm the tests pass on the current codebase, then confirm each correctness test actually fails when its criterion is violated (perturb the expected value or the input, watch it go red, then restore). A test that has never been seen to fail proves nothing.
4. Record tolerances and reference-data location — note the tolerance values and where reference/golden data lives. Cross-reference the ai-research-workflows:ensuring-reproducibility skill for test-data provenance (content hashes, retrieval dates, lockfile references).

Red flags — STOP, you're locking in a wrong result

The deadline temptation is to pin whatever the script prints and call it hardened:

Common Mistakes

• Exact float equality — using == on floating-point results fails on trivially different hardware or library versions; always use absolute or relative tolerances appropriate to the domain.
• Tests with no stated correctness criterion — a test that asserts a specific number without explaining why that number is correct is uninterpretable when it fails; document the analytical, reference, or published-result source for each expected value.
• Un-versioned reference data — storing golden outputs in a file without a version tag, content hash, or commit reference means a silent data change can break tests with no explanation; version reference data alongside the tests.
• Testing implementation details — asserting internal intermediate values rather than observable outputs couples tests to the implementation; test at the boundary where the correctness criterion applies.

Quality checklist

Before marking hardening complete:

• Correctness criteria are stated explicitly for each test
• Each baseline is verified against an independent reference, or labeled UNVERIFIED
• Each correctness test has been seen to fail when its criterion is violated
• All numerical comparisons use tolerances, not exact float equality
• Reference / golden data is versioned and its location is recorded
• Tests run in CI without manual intervention
• A failing test provides enough output to diagnose the problem

Cross-references

Integrates with the ai-research-workflows:validating-implementations skill (confirming the hardened code still satisfies its plan's success criteria) and the ai-research-workflows:ensuring-reproducibility skill (provenance for reference data and pinned baselines).