Ownership Boundary

Intelligence owns reviewable judgment over explicit inputs. It can rank candidates, apply declared policy, test counterfactuals, issue bounded recommendations, and refuse a decision. It cannot turn missing evidence into truth or convert a recommendation into an executable or laboratory action.

flowchart LR
    F["Foundation contracts"] --> I["Intelligence"]
    C["Core scientific results"] --> I
    K["Knowledge evidence records"] --> I
    I --> J["judgment and confidence"]
    I --> R["recommendation or refusal"]
    R --> L["Lab consequence review"]
    X["Runtime execution"] -. "not owned" .-> I

Owned Decisions

Decision surface	Representative implementation	Required evidence
candidate eligibility and ranking	candidates/filters.py, ranking.py, selection.py	explicit features, ranking policy, deterministic tie behavior
evidence posture	posture/evidence.py, posture/skeptical.py	named evidence references and uncertainty
benchmark judgment	judgment/benchmark_*.py	corpus, policy, blinded or counterfactual challenge, retained result
recommendation and refusal	judgment/recommendations.py, refusal.py	rationale, confidence, blocker, allowed next action
review packet construction	reviews/decision_briefs.py, external_review_kits.py	traceable inputs, policy version, challenge evidence
bounded learning	learning/adaptation.py	observed outcome record and governed adaptation rule

Refused Ownership

Concern	Canonical owner	Intelligence may consume
shared identifiers and provenance primitives	Foundation	typed contracts
parsing, normalization, quantification, statistics	Core	validated scientific results
claim custody, citations, lineage, contradictions	Knowledge	evidence records and graph queries
execution state, retries, replay, artifacts	Runtime	completed run contracts
intervention feasibility, follow-up, outcomes	Lab	consequence and outcome records

An Intelligence object must not silently redefine an upstream record merely because recommendation code needs a convenient shape. The active duplicate BeliefAuditEntry, BeliefAuditReport, and BeliefAuditSummary ownership finding demonstrates this exact risk and remains release-blocking.

Boundary Review

Accept a change here when the input meanings remain externally owned, the policy is explicit, the result carries rationale and uncertainty, and a refusal is possible. Relocate the change when it parses source evidence, controls a run, mutates evidence custody, or records laboratory truth.

The public contract is the decision transformation—not the authority of its inputs and not the downstream action taken from its result.