Skip to content

Decision policy configuration

Intelligence behavior is configured through typed, serializable policies—not global settings or environment variables. The policy is part of the evidence for a ranking or recommendation because changing weights, thresholds, metric direction, or tie-breaking can change the decision while inputs remain fixed.

Candidate ranking policy

RankingPolicy records:

  • stable policy_id, policy_family, and monotonic policy_version;
  • the minimum fraction and coverage of required metrics;
  • minimum evidence support and optional manufacturability floor;
  • weights for criteria, evidence, manufacturability, liability, and uncertainty;
  • uncertainty penalty, diversity bonus, and ordered tie-break rules;
  • a metric catalog defining class, unit, direction, and normalization.
from bijux_proteomics_intelligence.judgment.policies import (
    MetricDefinition,
    MetricDirection,
    RankingPolicy,
    ScientificMetricClass,
    ranking_policy_lineage,
    validate_factor_weights,
)

policy = RankingPolicy(
    policy_id="candidate-review",
    policy_version=3,
    minimum_evidence_support=0.65,
    metric_catalog=[
        MetricDefinition(
            metric_key="binding_kd_nm",
            metric_class=ScientificMetricClass.AFFINITY,
            unit="nM",
            direction=MetricDirection.LOWER_IS_BETTER,
        )
    ],
)

weight_audit = validate_factor_weights(policy)
lineage = ranking_policy_lineage(policy)

Do not run a governed ranking when required metric definitions are missing, factor weights are negative or unnormalized, or units and direction semantics are ambiguous. Preserve the policy fingerprint with ranking provenance.

Scenario and evidence gates

Progression, synthesis, scale-up, and redesign each have their own policy models. Their thresholds govern evidence support, confidence, blocker burden, residual risk, safety posture, and when a hold or redesign is required. evaluate_all_scenarios() keeps their actions separate before computing consensus.

Evidence readiness has independent configuration: assess_recommendation_readiness() accepts minimum trust and record-count thresholds, then also checks freshness, contradiction posture, evidence-kind diversity, decisive support, and source diversity. Passing a ranking threshold does not bypass this gate.

Configuration invariants

  • Candidate values and policy values are different inputs; never copy a score into a threshold to force a desired outcome.
  • A policy version change requires new lineage even when the identifier stays stable.
  • Missing evidence is not a zero-valued metric unless the metric contract says so explicitly.
  • Scenario disagreement and confidence spread remain report fields, not settings to suppress.
  • Promotion from advisory output to enforced policy is a recorded governance event, not a configuration default.