Evaluation and data quality
MetaFabric checks that data arrived complete and correct, and shows what needs fixing before anyone relies on it.
Two levels of checking#
A quality gate scores each record for completeness, validity and agreement between sources. It flags issues such as possible card numbers, and holds doubtful records back for review (see Unifying records).
A full evaluation in three gates. An operator, contributor or admin starts it from the Evaluation page or through MCP.
Three gates#
| Gate | What it checks |
|---|---|
| 1. Before the build | The ontology and mappings are valid and complete. |
| 2. After the build | Shape rules, record counts against the source, identity, data quality, test questions, and an AI review of meaning. |
| 3. Certification | Runs only when every blocking check passes and nothing is waiting for a person. |
Checks on mappings apply in every mode. Checks on records need the data brought in.
The completeness certificate#
When certification passes, MetaFabric issues a signed certificate covering the data, the change ledger and the versions used. Anyone in the workspace can verify it later in the console or through MCP.
Findings and repair#
Each failed check becomes a finding. The AI can suggest a repair, which changes nothing on its own. An approver, contributor or admin approves the fix before it is applied.
Approvals#
A mapping stops for approval before it publishes, the one step that cannot be undone. Every decision waiting on a person, from mappings, agents, findings and AI assistants, appears in one Approvals inbox.