Exploring schema evolution with ontology-driven propagation

A data engineering approach that uses a plain-English ontology as a runtime access policy to handle schema evolution automatically. Instead of maintaining a static column allowlist, the policy is written as natural-language rules (taxonomy + relationships), and an LLM applies those rules column-by-column using name patterns, data types, cardinality ratios, and value samples. The system handles ambiguous cases like high-cardinality text columns where names don't reveal PII — the LLM inspects sampled values to decide. Demonstrated on a fintech dataset with DuckDB and dlt, the approach correctly passed UUID-based user references while rejecting PII columns, with no code changes needed when new columns arrived. Limitations include numeric columns being treated as safe regardless of content, and no cross-column re-identification analysis.