Decision architecture

bijux-proteomics-intelligence makes analytical judgment reproducible without pretending judgment is evidence truth. It validates a candidate universe, applies a named policy, challenges the ranking, measures sensitivity and regret, then emits an advisory recommendation, downgrade, escalation, or refusal.

flowchart LR
    E["versioned evidence bundle"] --> C["decision context"]
    U["candidate universe"] --> C
    P["policy and constraints"] --> C
    C --> R["ranking"]
    R --> H["contradictions · falsifiers · scenarios"]
    H --> S["sensitivity · calibration · regret"]
    S --> O{"posture"}
    O --> A["recommend"]
    O --> D["downgrade or escalate"]
    O --> F["refuse"]

Responsibility map

Family	Owns	Retains for review
candidates	candidate models, validation, metrics, filtering, ranking, selection, lifecycle	complete universe, exclusions, fingerprints, score components
claims	claim-support interrogation	evaluated claim and immutable evidence references
interpretation	bounded analytical readings	assumptions, context, caveats, computed versus inferred content
judgment	policies, scenarios, recommendations, counterfactuals, sensitivity, confidence, regret	policy identity, alternatives, reversals, uncertainty
posture	evidence posture and skeptical review	downgrade, escalation, and human-review conditions
reviews	benchmark reviews, decision briefs, outsider packets, public scrutiny	input lineage, challenge results, unresolved pressure
learning	adaptation, refinement, convergence, stagnation	triggering outcomes and prior policy identity

The module map identifies concrete owners and dependency direction keeps evidence custody and laboratory authority outside the package.

Decision context is immutable input

sequenceDiagram
    participant K as Knowledge
    participant I as Intelligence
    participant L as Lab
    K->>I: evidence bundle at a fixed revision
    I->>I: validate, rank, challenge, calibrate
    I->>L: advisory record or refusal
    L-->>K: observation as new evidence
    K-->>I: new bundle for a new decision

Outcome learning produces a new calibration or policy record. It never edits the evidence bundle or recommendation that existed before the observation. State and persistence defines this history.

Ranking is policy, not fact

A ranking combines declared values: objectives, weights, constraints, thresholds, tie-breaking, missing-data treatment, and feasibility assumptions. Two valid policies may rank the same evidence differently. The architecture therefore preserves component scores and competing candidates rather than only the winner.

Execution model traces candidate intake through posture. The resulting record is explainable only when another reviewer can recompute the ordering from the same decision context.

Challenge before recommendation

Challenge is part of the decision path, not an optional report. Contradictions, falsifiers, blinded evidence, plausible alternative scenarios, threshold sensitivity, and regret test whether the ranking is stable enough for its declared use. An unstable result is downgraded, escalated, or refused even when its nominal score is high.

Extension rules

A new metric belongs with candidate quality; a new policy belongs with judgment; a new skeptical test belongs with challenge or posture; a new review artifact composes existing records without becoming an alternative policy engine. Every extension identifies its inputs, policy, deterministic behavior, failure modes, and authority limit.

Use extensibility model before adding a capability, integration seams for evidence and Lab handoffs, and error model for refusal and invalid decision contexts.

Architectural risks

The most serious risks are incomplete candidate universes, hidden exclusions, policy defaults that change without identity, scores presented without components, explanations generated from different inputs than the decision, and feedback that rewrites history. Duplicate decision models and broad owner facades also weaken accountability. Architecture risks and code navigation provide the focused review paths.