Repository fit

Intelligence owns the accountable transformation from reviewed evidence and constraints into an advisory decision. Its output is not “the answer”; it is a versioned record of which candidates were considered, which policy ranked them, how the ordering responded to challenge, and why the result was recommended, downgraded, escalated, or refused.

Why a separate package exists

Evidence and judgment must be independently reviewable. If ranking lives in Knowledge, the evidence owner can silently determine action. If ranking lives in Runtime, execution convenience can become policy. If it lives in Lab, capacity and operational burden can be mistaken for scientific support.

flowchart LR
    evidence["versioned evidence"] --> policy["named decision policy"]
    candidates["candidate universe and exclusions"] --> policy
    constraints["values, feasibility, and objectives"] --> policy
    policy --> challenge["scenarios · sensitivity · falsifiers · regret"]
    challenge --> disposition{"defensible?"}
    disposition -->|yes| recommend["bounded recommendation"]
    disposition -->|unstable| escalate["downgrade or human review"]
    disposition -->|no| refuse["refusal with unmet conditions"]

Owned surfaces

Surface	Intelligence responsibility
candidates	candidate identity, validation, exclusions, quality, lifecycle, ranking, and selection
interpretation	policy-facing readings of governed scientific results
claims and skeptical review	expose support requirements, contradictions, and falsifiers to judgment
judgment	policy, scenarios, counterfactuals, sensitivity, confidence, regret, recommendation, and refusal
posture and reviews	declare evidence posture and assemble challengeable decision packets
learning	create new calibration or policy records from outcomes without rewriting history

Placement test

The package owns a rule when changing the decision values or constraints may change the result while the underlying scientific and evidence records remain unchanged.

Change	Attribution
scientific metric or result changed	Core
source, contradiction, or evidence context changed	Knowledge
weights, hard constraints, objective, scenario, or tolerance changed	Intelligence
execution provider or artifact changed	Runtime
feasibility, cost, capacity, or observed outcome changed	Lab, then a new Intelligence review

What does not fit

• hidden scores without a candidate universe, policy identity, or explanation;
• models that rewrite evidence strength as part of ranking;
• execution or laboratory automation disguised as a “next step”;
• confidence detached from sensitivity, alternatives, and evidence posture;
• output that cannot refuse when hard constraints or support are inadequate;
• outcome-aware learning that edits the historical decision instead of creating a linked successor.

Fit tests

A decision feature belongs here only when a reviewer can reproduce its inputs, policy, ordering, challenge results, and disposition without granting the package authority over evidence or action. The human-review state remains explicit, and Lab performs its own readiness assessment before any executable handoff.

Continue with recommendation record anatomy, compare decisions, and known limitations.