This corpus mixes brevo with deterministic documents balancing. Training and validation rows only are included; source test rows remain excluded. Exact source manifests are retained under provenance/, and source-aware validation can replay every selected row against the authoritative datasets.
