Declare a data domain fit for use in a production model
Draft
Declare a data domain fit for use in a production model
Allocation
| L3-03 | |
|---|---|
| Decides | Data Steward |
| Consulted | Chief Data Officer and Model Owner |
| Executes | Data Engineering |
| Evidence | Data quality report, lineage record, classification |
In plain terms
Judge whether this data is accurate, complete and traceable enough that a production decision may rest on it. Fit for one purpose is not fit for all purposes.
What is being judged
Fitness for a stated use, which is why the declaration names the model or the class of use rather than blessing the domain generally.
Four things carry it. Accuracy against reality, not internal consistency; a domain can be perfectly consistent and systematically wrong. Completeness for the population the model serves, since a domain complete for existing customers may be sparse for the segment the model is meant to reach. Lineage, meaning the path from source to model input is traceable, because an untraceable input cannot be corrected when it turns out to be wrong. Classification, which determines what controls the data carries into the model.
What this decision does not cover
It does not define the master record, which is L3-05 and belongs to the Chief Data Officer. It does not approve access, which is L3-06. It does not determine lawfulness, which is L4-REG-03.
When it fires
On event. Before a production model first consumes the domain. On material change to the source, schema or collection method. On a model purpose extension under L4-AUT-04. At the stated review date. On a data quality incident.
On cycle. Per domain review cycle.
What you need before deciding
The data quality report with actual measures rather than a process description. Lineage from source to input. Classification. Which model and which purpose, since the declaration is against a use. Known gaps, which is the part most often omitted and most useful.
How this goes wrong
Declaring the domain rather than the use: a general fitness statement that a second model then relies on for a purpose the steward never considered. Fitness by pipeline existence: the data arrives through a working pipeline, therefore it is fit, which conflates delivery with quality. The unstated gap: a declaration that records what is good and omits what is missing, so a Model Owner cannot judge whether the gap matters for their case.
Related decisions
Parallel L2-05 and L3-04 interface fitness, the same judgment applied to data in motion.
Upstream L3-05 master record definition.
Downstream L3-01 release to validation, L4-AUT-01 authorization.
Related L3-12 feature reuse, which carries fitness into a second context.
Instrument references
EU AI Act Article 10 requires training, validation and testing data sets be relevant, representative and to the extent possible free of errors. ISO/IEC 42001 Annex A data controls. DAMA-DMBOK covers data quality practice. None requires a per-use declaration by a named steward.
Correction
The maintainer answers corrections. There is no service level. Responses are best-effort and opportunistic within a reasonable time: a correction raised on a Monday is answered that week or sooner.