Analyst
Builds dashboards in Looker, Tableau, or Power BI. Trusts upstream data by convention — if the numbers look reasonable, ship the report.
Every input carries a SHA-256 seal from ingest. The analyst does not need to trust the pipeline — the seal is verifiable, independently, at any time.
Engineer
Writes ETL/ELT jobs. Validates with row counts, null checks, and schema tests. No cryptographic chain ties input to output.
Config seal applied before the first data touch. Any re-run with different parameters produces a different seal — detectable, not deniable.
Data scientist
Trains on "last known good" data. No provenance seal on the training set. Model lineage starts at the notebook, not at the source.
Cross-grain audit: SUM(grain totals) = grand total. Catches what row-count tests miss. Training data provenance is part of the sealed chain.
Pipeline CI/CD
Tests pass or fail on logic correctness, not on data integrity across environments. A test that passes on dev may silently diverge on prod.
Medallion chain holds across local, container, and cloud. Same data, same config, same seal — at every compute tier. Cross-tier parity is a test, not a claim.
AI agent
Reads data from whatever endpoint is configured. No mechanism to verify the pipeline was not tampered with between training and inference.
Same seal mechanism works for agent-sourced data. Agent identity and data provenance travel in the same chain. An agent that cannot prove its data source cannot act on it.
Audit
Manual attestation. Screenshots. Trust the person who ran the job. The evidence is a report about a process, not the process itself.
The artifact is the evidence. No attestation theater. The auditor receives the sealed chain — not a report about a sealed chain.