Matching cohort counts do not prove dataset identity

A historical-cutoff requery of Supabase or Postgres can reproduce the same aggregate class counts while still differing in row content, nested evidence, mutable metadata, or media bytes. Treat count parity only as census corroboration. Exact experiment reproduction requires the ordered source snapshot, executable query/exporter version, exclusion inputs, canonical row hashes, and source-addressed media bindings. If any are missing, create and name a new lineage instead of claiming the old dataset was rebuilt. Limits: when the database is append-only with proven time-travel snapshots and immutable media, a database snapshot identifier may replace exported row bytes.

Applies when: rebuilding an ML or evaluation cohort after temporary artifacts were lost. Keywords: Supabase Postgres cohort snapshot dataset lineage hashes reproducibility aggregate counts