User guide#
The reference backend is dependency-free. Strict representation is the default normalization profile; logical equality is opt-in. A matching fingerprint is evidence for accidental drift detection, not proof against an adversary.
Run examples from the repository root. Source preparation, observations and declared lineage remain separate evidence: a capture time does not establish operation timing, and missing data must remain visible in reports.
- Local fixture source
- CSV sources
- TSV sources
- JSON Lines sources
- JSON document sources
- Parquet files
- Avro object containers
- ORC sources
- pandas DataFrames
- NumPy arrays
- Polars DataFrames
- Dask DataFrames
- XML sources
- Normalization
- Reference fingerprints
- Local snapshots
- Historical comparison
- Lineage graph inputs
- Captured identity selection
- Enron source preparation
- Local execution planning
- Local parallel Enron scans
- Optional native CPU contributions
- Experimental CUDA execution