Decision foundations

Intelligence owns the policy that transforms reviewed evidence and program constraints into a ranked action, downgrade, escalation, or refusal. Its output is an inspectable judgment under a declared policy. It is never a replacement for the scientific result, evidence record, execution history, or laboratory decision on which it depends.

flowchart LR
    E["immutable evidence references"] --> J["decision context"]
    C["program constraints"] --> J
    P["declared policy"] --> J
    J --> R["ranking"]
    R --> H["challenge"]
    H --> S["sensitivity and regret"]
    S --> O{"posture"}
    O -->|stable| A["advisory recommendation"]
    O -->|uncertain| D["downgrade or escalate"]
    O -->|unsupported| F["refusal"]

Decision routes

Question	Guide	Governing record
What does Intelligence own?	Package overview	decision context and policy output
Is a proposal policy or upstream truth?	Ownership boundary	canonical owner and immutable input reference
Which analytical surfaces exist?	Capability map	candidates, interpretation, challenge, judgment, posture, reviews
What is outside the decision boundary?	Scope and non-goals	scientific calculation, evidence truth, execution, and lab authority
Does the recommendation survive pressure?	Workflow recommendation challenges	blinded and counterfactual challenge record
Is confidence calibrated?	Workflow recommendation confidence	sensitivity, calibration, and regret record

This package does not own resolves common category errors, including copying evidence into a policy model or presenting a recommendation as laboratory authorization.

Fact, policy, and action

Layer	Owns	Intelligence treatment
Core	scientific result and benchmark acceptance	reference as an input; never recalculate silently
Runtime	run identity, provider, state, and artifacts	use execution evidence; never rewrite run history
Knowledge	claims, sources, context, and contradictions	consume a versioned review bundle; never edit evidence truth
Intelligence	ranking, challenge, sensitivity, posture, and refusal	create a new decision record under a named policy
Lab	readiness, execution authority, observation, and QC	hand off an advisory action; receive outcomes as new evidence

This separation allows two policies to reach different rankings from the same evidence without pretending that the evidence itself changed.

Decision protocol

1. Validate and fingerprint the complete candidate universe, including explicit exclusions.
2. Resolve immutable references to scientific, execution, and evidence inputs.
3. Apply a named policy with visible constraints, weights, thresholds, and tie-breaking.
4. Preserve score components and competing candidates, not only the winner.
5. Challenge the ranking with contradictions, falsifiers, blinded evidence, plausible scenarios, and counterfactuals.
6. Measure sensitivity, calibration, and regret.
7. Emit a recommendation, downgrade, escalation, or refusal with reason codes and human-review posture.

sequenceDiagram
    participant K as Knowledge
    participant I as Intelligence
    participant L as Lab
    K->>I: versioned evidence review bundle
    I->>I: rank, challenge, test sensitivity
    I->>L: advisory record or refusal
    L-->>K: observed outcome with provenance
    K-->>I: new evidence version for a new decision

The feedback loop creates a new recommendation. Historical decisions remain attached to the evidence and policy available when they were made.

Challenge and confidence

The recommendation challenge route asks what evidence pattern would reverse or weaken the ranking. The recommendation confidence route asks whether expressed confidence matches observed and benchmark behavior. Together they expose:

• dependence on one fragile feature, threshold, or source;
• hidden alternatives that become preferable under plausible constraints;
• overconfidence, underconfidence, and poorly calibrated refusal;
• regret when a different action would have produced a better consequence;
• workflow-family ceilings inherited from upstream evidence.

Strong explanation without these checks is still only persuasive prose.

Evolution rules

The domain language stabilizes terms such as candidate, policy, scenario, falsifier, posture, confidence, and regret. The change principles require policy identity and comparison when ranking behavior changes. Dependencies and adjacencies and repository fit preserve one-way ownership, while the lifecycle overview connects candidate intake through decision and downstream outcome without granting Intelligence execution or lab authority.