Document simulation work in four layers so research evidence is not mistaken for production capability

For an evolving simulation subsystem, keep four distinct documentation layers: a normative ontology/specification, a current live-boundary report, a standalone evaluator report, and machine-verifiable provenance manifests. The evaluator report must explicitly separate causal state from removable readouts and decorative assets. The live report should say what is wired, what remains isolated, and why. Store exact run outputs and hashes in generated summaries/manifests; if prose includes test counts or verdicts, label them as dated snapshots because they become stale quickly. Preserve the exact generation prompt when generated art is part of an evidence bundle, even if that art is decorative only.