Defensive coding for research software

Research code's worst bug is the one that runs to completion: the unit mismatch, the silently dropped rows, the column shifted by one, the reused random stream - producing plausible, wrong numbers that reach a paper. Defensive coding is the discipline of making wrongness LOUD: validate at boundaries, make implicit physics and randomness explicit, and prefer a crash today to a correction notice next year. It complements rseng-numerical-accuracy (float behavior) and feeds rseng-testing (the checks become tests).

Validate at the boundaries

Data enters code at boundaries - files, instruments, APIs, user parameters - and every boundary is a corruption opportunity:

• Schema-check tabular data on ingest: expected columns, dtypes, units, ranges, allowed categories, uniqueness of keys (pandera-style declarative schemas for DataFrames; JSON Schema for configs and records). The schema is executable documentation of the data contract (rseng-data-management's data dictionary, enforced).
• Sanity-check the science, not just the types: physical ranges (no negative concentrations, latitudes within +/-90), plausible magnitudes, conservation totals, expected row counts within tolerance of yesterday's. Domain assertions catch what dtype checks cannot.
• Missing data is a decision, not a default: silent NaN propagation and silent row-dropping (the pandas default in many operations) are the classic silent killers - count and REPORT what was excluded and why, and make exclusion rules explicit code (rseng-research-integrity checks this at the manuscript end; this skill prevents it at the source).
• Fail loud, fail early: raise on contract violations at ingest rather than letting bad data flow downstream; in batch settings quarantine-and-log per item (rseng-big-data-processing) - but never silently skip.

Units and quantities

Unit errors are the canonical silent research bug - flagship missions have died of them:

• Attach units in code, not in comments: a quantity library (pint; astropy.units in astronomy) makes units part of the value, checked at every operation - adding meters to seconds raises instead of publishing.
• Enforce at boundaries even when the core stays plain-numeric for performance: convert-and-strip on ingest (asserting the expected unit), reattach on output, and document the internal convention in ONE place (rseng-scientific-file-formats' metadata discipline carries units in the files themselves).
• Degrees vs radians, per-second vs per-minute, and log-vs-linear are unit bugs in spirit: name them in variable names or types when a full quantity system is overkill (angle_rad, rate_per_s).

Disciplined randomness

• Explicit generators, never global state: create a seeded RNG object (numpy's Generator API) and pass it - seeding the global makes order-dependent, library-colliding randomness (the Scientific Python RNG guidance is the reference).
• Parallel work gets derived streams: spawn per-worker generators from a master seed (seed sequences), never the same seed in every worker - identical streams across "independent" replicates is a silent statistics-destroyer (rseng-hpc-computing job arrays included).
• Seeds are provenance: record them with outputs and in configs (rseng-reproducibility owns the bookkeeping); stochastic tests assert distributions or use fixed seeds knowingly (rseng-testing).

The habit, proportioned

Not every script needs schemas: exploration code needs only the cheap habits (explicit NaN policy, named units, seeded RNGs); shared pipelines add ingest validation; published analyses add the full contract checks in CI (rseng-ci-cd) so drift in upstream data sources is caught at the pull request, not in the plot. When a defensive check fires in production, keep it as a regression test - each catch is a documented failure mode (rseng-trainer's errors-are-curriculum applies: explain what the check just prevented).

Working with this skill

This skill is source-independent: its authority is the tool documentation and Scientific Python guidance linked below. It is the prevention layer for the failure modes rseng-research-integrity hunts post-hoc.

Learn more (verified):

• https://pandera.readthedocs.io - pandera DataFrame validation
• https://pint.readthedocs.io - pint physical quantities
• https://json-schema.org - JSON Schema for configs and records
• https://docs.astropy.org/en/stable/ - astropy (units module)
• https://blog.scientific-python.org/numpy/numpy-rng/ - Scientific Python guidance on NumPy random number generators

Related skills

Check whether any of these applies before moving on:

• rseng-big-data-processing - quarantine-and-log at batch scale
• rseng-data-management - data dictionaries the schemas enforce
• rseng-hpc-computing - per-worker RNG streams in parallel jobs
• rseng-numerical-accuracy - float behavior behind silent errors
• rseng-research-integrity - post-hoc hunt for same failures
• rseng-testing - fired checks become regression tests