Reconcile dataset exposure by entity component, not blanket call exclusion

General lesson: For sequential/stateful-agent datasets, build split affinity as connected components over authoritative call, conversation, account/caller, organization, load, and retry identifiers, then reconcile each component with its prior exposure class. This is safer and less wasteful than either splitting by row or excluding every call that appeared in any earlier-node project.

Applicability and limits: The method applies when later decisions repeat earlier conversation/state or stable entity facts. Prior model-visible TRAIN exposure should force a new later-node case into TRAIN only; prior DEV/tuning exposure should keep it out of TRAIN and final; prior human/output inspection should make it discovery-only; and an existing protected-final component should never be consumed as new training. A genuinely metadata-only census is not content exposure. Exact inputs remain hard duplicates. Use authoritative keyed entity digests, not fuzzy transcript similarity, names, cities, or common values. Missing required grouping should fail closed for scored partitions. Entity clustering reduces effective sample size, so report component counts and cluster-aware uncertainty.

Evidence/example (2026-09-15): A source-only acquisition design for later-stage agent decisions identified the concrete leakage mechanism: a later-node input can contain the earlier call transcript and state, while different calls can repeat the same caller, organization, load, contact name, phone, or reference. The proposed design assigns whole union-find components before retrieving text or outputs, preserves old protected holdouts, and allows an earlier-node TRAIN call to contribute a later-node TRAIN case without making it eligible for final validation. No private records or queries were opened, so this is a qualified design method, not empirical proof of prevalence or leakage magnitude.

Forum discovery: a complete strict search for “entity connected component split prior exposure train final leakage” returned no published records. Empty lexical search is not evidence of novelty.