Test strategy

Knowledge testing applies pressure to evidence custody: incomplete sources, ambiguous identities, duplicated lineage, contradiction, stale records, unresolved review, serialization, and downstream interpretation.

Evidence layers

Layer	Contract under test	Representative suite
reference custody	registry integrity, source identity, citation, license and freshness fields	tests/references/
biological identity	exact, ambiguous, absent, obsolete, cross-species, isoform, and namespace mapping	identity and biological namespace tests
claim and evidence models	valid/invalid construction, context, provenance, confidence meaning, and immutability	tests/memory/test_claims.py, test_evidence_bundle.py
graph integrity	endpoints, edge types, support, contradiction, orphan prevention, and round trip	test_evidence_graph.py
reconciliation	competing contexts, deterministic policy, hold, unresolved, and audit trace	test_resolution.py and contradiction surfaces
review state	fixed revision, complete adverse evidence, disposition, rationale, and stable assembly	tests/reviews/
grounding	citations, benchmark lineage, literature coverage, deficits, contradiction, and release ceiling	workflow grounding and literature tests
package boundary	Foundation serialization, Core results, Intelligence references, and Lab feedback remain aligned	tests/package/ and consumer tests

Challenge route

flowchart LR
    R["source and identity"] --> M["claim and evidence models"]
    M --> G["graph integrity"]
    G --> C["contradiction and reconciliation"]
    C --> V["review revision"]
    V --> B["consumer bundle"]

Run the focused source, memory, review, or grounding suite first, then the full package suite for shared models, persistence, registries, and public outputs:

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

Required imperfect evidence

Include missing sources, duplicate lineage, ambiguous identifiers, stale records, unsupported namespaces, conflicting claims, context-dependent agreement, unresolved reconciliation, and incomplete review. An idealized single-source graph cannot establish honest uncertainty handling.

Fixture replay proves deterministic handling of retained records. Live-source availability and freshness require separate retrieval evidence. Keep those claims distinct in test reports.