Integration Seams

Intelligence is the advisory layer between scientific evidence and accountable review. It consumes typed analysis and curated context, makes policy-driven judgment explicit, and emits recommendations that downstream lab systems may consider but must never execute implicitly.

flowchart LR
    C[Core scientific artifacts] --> E[Candidate and claim framing]
    K[Knowledge evidence and references] --> E
    E --> P[Evidence posture and contradictions]
    P --> J[Policy and scenario judgment]
    J --> B[Decision brief and review packet]
    B --> L[Lab feasibility and accountable approval]
    B --> R[Runtime persistence and transport]
    L --> O[Observed outcome]
    O --> A[Prospective learning]
    A --> J

Seam obligations

Seam	Upstream obligation	Intelligence obligation
Core → intelligence	provide valid scientific models, design and QC state, claims, uncertainty, and provenance	evaluate without renormalizing measurements or inventing missing scientific support
Knowledge → intelligence	provide evidence graph, source context, trust, freshness, contradiction, and caveat state	preserve identifiers and adverse evidence; do not curate sources locally
Intelligence → review	provide policy identity, factor contributions, rankings, scenario outcomes, refusals, falsifiers, uncertainty, and alternatives	reviewer records accountable disposition and rationale
Intelligence → lab	provide a bounded recommendation and named evidence gaps	lab re-evaluates assay risk, controls, materials, capacity, and execution authority
Runtime ↔ intelligence	transport typed inputs and persist outputs with lineage	keep judgment deterministic and free of workspace, provider, retry, and credential concerns
Lab outcome → learning	provide attributable observations and outcomes	adapt future posture without rewriting historical decisions

Authority boundary

An advance or scale_up scenario result is an analytical recommendation, not authorization. Readiness scores, confidence, and candidate rank cannot stand in for a review-board decision, a lab handoff, or local safety approval. The lab package owns the operational translation and can refuse a recommendation that is scientifically interesting but infeasible or unsafe.

Learning is similarly one-directional in time. An outcome can inform a new policy or refinement run, but the package must retain the policy, evidence, and rationale used for the earlier decision. This separation makes regret and drift measurable instead of silently revising history.

Contract change impact

When core changes a claim or metric, review candidate validation, claim support, refusal thresholds, interpretation, scenario coverage, and benchmark packets. When knowledge changes evidence state, review freshness, contradictions, belief audits, recommendations, and decision briefs. When intelligence changes a policy, fingerprint and compare rankings, scenarios, counterfactuals, sensitivity, and review outputs before downstream consumers adopt it.