Test strategy

Intelligence testing proves both the decision and its explanation under ordinary, adverse, ambiguous, and changing conditions. Numeric movement alone is not sufficient evidence.

Evidence layers

Layer	Question	Representative suite
candidate integrity	are valid, invalid, missing, duplicate, excluded, and fingerprinted candidates handled explicitly?	tests/candidates/
component semantics	do orientation, scale, boundaries, missingness, and explanations agree?	candidate ranking and quality tests
policy behavior	are constraints, weights, thresholds, ties, alternatives, and order deterministic?	judgment policy and decision tests
challenge	do contradictions, falsifiers, counterfactuals, blinded cases, and skeptical posture affect the result?	contradictions, falsifiers, judgment challenge, and posture tests
stability	which plausible evidence, threshold, weight, or scenario changes reverse or weaken the action?	sensitivity and scenario tests
calibration and regret	does declared confidence match the corpus and are alternative costs retained?	confidence, calibration, and regret tests
review artifact	can a reader reconstruct evidence, candidates, policy, challenges, posture, and authority?	tests/reviews/
learning	does outcome feedback create a versioned policy without rewriting history?	tests/learning/
package boundary	are Foundation, Core, Knowledge, Runtime, and Lab meanings preserved?	tests/package/

Decision challenge path

flowchart LR
    F["fixed evidence and candidates"] --> P["policy result"]
    P --> C["contradiction and counterfactual"]
    C --> S["sensitivity"]
    S --> G["calibration and regret"]
    G --> R["review artifact and posture"]

Run the closest decision family first, then the complete suite when public policy, candidate models, review artifacts, or cross-package contracts move:

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

Required negative outcomes

Test no candidate, tied candidates, contradictory evidence, missing decision-critical evidence, unstable recommendation, calibration outside the supported corpus, excessive regret, and absent downstream authority. Hold and refusal are successful test outcomes when their preconditions are met.

Snapshotting a recommendation sentence does not prove the policy. Assert the candidate set, component results, alternatives, challenge findings, posture, and authority record that explain it.