Synthetic Data, Defects & Provenance
Authoritative formatted source: Google Doc · PDF
Design principles
Meridian source data is intentionally synthetic and intentionally imperfect. Defects are learning assets, not accidental corruption.
- No real PII, employer data, credentials, or production extracts.
- Use deterministic generation when exact expected results matter.
- Preserve canonical raw sources.
- Record intentional defect IDs and their expected manifestations.
- Separate source data from staged/curated/repaired outputs.
- Keep story scale and lab scale distinct.
- Give every file/table a business-process and source-system context.
Canonical starter data
The blueprint includes small, inspectable source sets such as:
locations.csvcustomers_crm.csvcustomers_pos.csvcustomers_service.jsonproducts.csvsuppliers.csvorders.csv/ relational ordersorder_items.csv/ relational order itemsinventory_snapshot.csvreference_codes.csvdocuments_registry.csv- defect map / validation summary / source manifest / README
Later chapters may generate larger JSON/Parquet event data when the learning objective requires it.
Defect strategy
Intentional defects should map to known Meridian problems, such as:
- missing or stale customer contact data;
- duplicate identity candidates;
- invalid product/reference relationships;
- inconsistent category/reference codes;
- impossible or suspicious quantities;
- source-format drift;
- incomplete metadata;
- access exposure in controlled security labs.
Do not “clean up” the raw source and erase the defect. Repair should produce a new staged/curated/master/evidence state or use a reproducible defect-injection mechanism.
Provenance record
For each generated or supplied dataset, preserve enough information to reproduce it:
| Field | Example of what to record |
|---|---|
| Dataset/file/table | Exact name |
| Generator/source | Script or source manifest |
| Generation parameters | Seed/row-count/options where applicable |
| Business process | Browse-to-Order, Service, etc. |
| Source system | CRM, POS, SERV, PIM, OMS, etc. |
| Intentional defects | Stable defect IDs / descriptions |
| Expected validation | Row count, uniqueness, reference checks, known exceptions |
| Canonical raw location | data/raw/... or Drive source folder |
| Derived outputs | Staging/curated/evidence locations |
Important seed note
The blueprint demonstrates deterministic generation using explicit generation parameters. The actual baseline seed/value should be taken from the live source manifest or generator configuration for the dataset being used, rather than assumed from an example in a guide. This prevents a documentation example from becoming an accidental false source of truth.
Validation before lab use
Before beginning a lab that depends on exact source behavior:
- Verify files/tables are present.
- Check expected row counts.
- Check primary uniqueness/reference expectations where applicable.
- Confirm the intended defect set is present.
- Confirm raw files have not been overwritten by a prior exercise.
- Record the run/source version in evidence when exact reproducibility matters.
Source: Meridian Synthetic Data & Defect Blueprint.