Local development

Intelligence turns governed evidence into advisory rankings, challenges, scenarios, refusals, and review packets. A local change is safe only when a reviewer can recover which input, policy, or uncertainty changed the advice. Higher scores or more decisive outputs are not success criteria by themselves.

Run the package gates

Use package dispatch from the repository root:

make lint PACKAGE=bijux-proteomics-intelligence
make test PACKAGE=bijux-proteomics-intelligence
make quality PACKAGE=bijux-proteomics-intelligence
make api PACKAGE=bijux-proteomics-intelligence

Run make build PACKAGE=bijux-proteomics-intelligence when public modules, compatibility exports, metadata, or packaged fixtures change. Generated results remain under artifacts/.

Isolate the decision lever

flowchart LR
    evidence["versioned evidence"]
    policy["named policy and thresholds"]
    analysis["ranking · challenge · scenarios"]
    advisory["advice or refusal"]
    review["human review packet"]
    evidence --> analysis
    policy --> analysis
    analysis --> advisory --> review

Change one decision lever at a time: factor direction, weight, hard constraint, threshold, tie policy, refusal rule, or scenario assumption. Keep the previous policy fixture available and compare the complete output, not only the final action. Candidate exclusions, factor contributions, Pareto state, contradictions, falsifiers, confidence, and unresolved questions can reveal meaningful drift hidden by an unchanged recommendation.

Choose evidence by change type

Change	Required evidence
ranking policy	factor-level reasons, ties, exclusions, and order stability
refusal rule	accepted boundary, refused boundary, and required remediation
scenario logic	agreement, disagreement, hold pressure, and escalation cases
counterfactual	named perturbation and whether the action changes
report contract	all decision inputs and policy lineage remain recoverable
learning update	prior decision is immutable and new outcome linkage is explicit

Use synthetic fixtures to isolate policy behavior and representative governed fixtures to establish integration. Label synthetic evidence honestly; it must not appear to be an observed scientific result.

Preserve authority boundaries

Intelligence remains advisory unless an explicit promotion record gives an output enforced status. It does not execute workflows, curate source evidence, or authorize laboratory work. Refusal is a first-class result when evidence, design, or review requirements are insufficient.

The change is ready when identical inputs and policy remain deterministic, each changed recommendation has an inspectable reason, uncertainty is not erased, refusal and human authority remain explicit, and review artifacts reconstruct the decision without private implementation knowledge.