Keep source and conclusion separate
A paper, dataset, simulation, and agent interpretation are different evidence types. Preserve their relationships rather than collapsing everything into a confidence score.
Expose assumptions
List model boundaries, missing measurements, conflicting sources, and conditions under which a conclusion could fail. Summarize decision rationale without exposing hidden model deliberations.
Close with accountable approval
The final package should make the decision owner, version, approval state, and unresolved risks clear. This is the foundation of a reviewable handoff.