Test strategy

Core testing selects evidence from the scientific claim. Unit coverage protects local rules; public trust additionally needs malformed inputs, reference or property evidence, benchmark challenge, provenance, and transfer limits.

Evidence ladder

Layer	Purpose	Typical evidence
contract	construction, types, schema, reason codes, and lifecycle	domain and public-model tests
algorithm	equations, orientation, boundary values, invariants, and deterministic output	unit, parametrized, and property tests
adversarial input	malformed formats, contradictions, missingness, decoys, contaminants, and unsupported cases	parser mutation and negative-science corpora
reference	agreement with an independently specified value or implementation	reference fixtures and tolerance tests
regression	preserved behavior for a previously observed failure	minimal named fixture with expected disposition
benchmark	primary package, companion pressure, acceptance bars, provenance, and artifact review	workflow-family benchmark suites
transfer	behavior under changed cohort, matrix, library, engine, or design	holdout and generalization surfaces
boundary	schema, CLI/API parity, Runtime handoff, and downstream interpretation	package and cross-package tests

Select evidence by change

flowchart TD
    C["changed scientific claim"] --> K{"kind"}
    K -->|parser| P["formats, provenance, malformed and rejected rows"]
    K -->|algorithm| A["reference, properties, boundaries, sensitivity"]
    K -->|workflow| W["primary, companion, negative, transfer, acceptance"]
    K -->|artifact| R["schema, round trip, lineage, consumer"]
    P --> B["boundary and regression proof"]
    A --> B
    W --> B
    R --> B

Run the focused domain or workflow suite first, then the complete Core suite when shared models, public APIs, workflow posture, or cross-cutting helpers move:

uv run --project packages/bijux-proteomics-core \
  pytest -q packages/bijux-proteomics-core/tests

Independence and circular proof

A benchmark generated by the changed algorithm cannot independently validate that algorithm. A parity test that compares two adapters over the same imported producer output establishes adapter agreement, not independent scientific truth. Record the origin of expected values and separate reference evidence from regression snapshots.

Performance and determinism

For chunking, parallelism, vectorization, or caching, pair resource evidence with serial equivalence, stable ordering, partial-failure behavior, and artifact identity. Faster output that changes accepted records or provenance is a scientific regression.