Decision interfaces

Intelligence interfaces expose candidate evaluation, ranking, skeptical challenge, recommendation posture, and review artifacts. Their contract is to preserve how a decision was produced: the candidate universe, evidence revision, policy, score components, alternatives, sensitivity, and refusal or human-review conditions.

flowchart LR
    E["evidence and scientific records"] --> D["decision context"]
    D --> R["ranking result"]
    R --> C["challenge record"]
    C --> P["posture"]
    P --> B["decision brief"]
    B --> H["human or Lab handoff"]

Choose an owner interface

Need	Owner module	Expected output
validate or compare candidates	candidates	validated set, exclusions, quality and fingerprints
rank or select candidates	candidates.ranking and selection owners	component scores, ordering, ties, policy identity
interrogate claim support	claims, contradictions, falsifiers	support and challenge findings with evidence references
interpret scientific results	interpretation	bounded interpretation with assumptions and caveats
evaluate scenarios or recommendations	judgment	scenario, counterfactual, sensitivity, confidence, regret, recommendation
determine evidence posture	posture and refusal	advisory, downgrade, escalation, hold, or refusal reasons
assemble external review	reviews	report contract, decision brief, outsider packet, scrutiny record
propose follow-up	next_steps and learning	evidence request or policy-learning record, not execution authority

The package root exposes 14 lazy-loaded owner modules:

from bijux_proteomics_intelligence import candidates, judgment, reviews

Import domain symbols from their owner module. API surface maps supported paths and public imports distinguishes the candidate representations exposed for different responsibilities.

Decision record

A portable recommendation or refusal identifies:

• decision and candidate-set identity;
• the exact Core, Runtime, and Knowledge artifact references evaluated;
• candidate exclusions and validation findings;
• policy name, version, objectives, constraints, thresholds, and tie-breaking;
• score components, ordering, alternatives, and Pareto relationships;
• contradictions, falsifiers, scenarios, and counterfactual results;
• sensitivity, calibration, confidence, and expected regret;
• advisory, downgrade, escalation, hold, refusal, and human-review posture;
• proposed next evidence without implying permission to execute it.

Data contracts defines these semantics and artifact contracts defines portable review forms.

Questions and explanations

Query and interpretation interfaces answer questions against a fixed decision context. Explanation reports the declared reasoning path; it must not invent missing evidence or conceal policy-sensitive alternatives.

flowchart TD
    Q["review question"] --> X["resolve decision context"]
    X --> E["retrieve evidence and policy"]
    E --> A["answer with lineage"]
    E --> U["unresolved or unsupported"]

An unsupported question produces an explicit limitation or refusal. It is not answered by extrapolating beyond the available evidence.

Configuration and compatibility

Weights, objectives, constraints, thresholds, missing-value handling, tie-breaking, challenge sets, and posture rules are public decision configuration. Their normalized identity belongs in every result. See configuration surface.

Candidate fields, policy identifiers, score orientation, posture enums, reason codes, artifact schemas, and default behavior are compatibility surfaces. Compatibility commitments defines comparison and migration requirements when they change.

Authority boundary

Intelligence can advise, challenge, request evidence, or refuse. It does not rewrite Knowledge provenance, execute Runtime work, authorize candidate progression, or approve laboratory activity. Operator workflows shows how those handoffs remain explicit.