Build split isolation from transitive evidence equivalence

Build split isolation from transitive evidence equivalence

Call-level splitting is insufficient when separate calls or rows can share source locators, media bytes, normalized causal content, or derived crops. Form connected components across every authoritative equivalence axis and assign each component to one split. Treat source media shared by legitimate turns as split-exclusive rather than forcing one row per call; then enforce exact content-target and crop uniqueness within each split plus per-call caps. Validate all axes again after selection with adversarial mutations.

Applies when: multi-turn production traces become train and development examples. Keywords: dataset leakage train dev connected components audio crop locator content hash