Engineering, production, governance, cost, communication · insight-1304
Hand engineers a specification with test vectors, not a notebook
- Evidence type
- practitioner heuristic
- Difficulty
- medium
- Roles
- Data scientist, ML engineer
- Industries
- cross-industry
Summary: situation, action, reason, boundary
Situation
A data scientist passes a notebook to engineers to reimplement in the serving stack (hypothetical).
Action
Deliver the input schema with units and null rules, the transformation and scoring specification, pinned library versions, and a file of input and expected-output pairs including edge cases with agreed tolerances. Engineers run the vectors in their CI.
Reason
A file of executable examples turns a prose description into a checkable contract and exposes mismatches before traffic does.
Boundary
Any model whose logic is reimplemented by another team or language.
Symptom or risk
The reimplementation differs in rounding, missing-value handling, category order and time zones, and each team blames the other when production disagrees with the notebook.
Exceptions
If the exact artefact is deployed as is, test vectors still guard environment changes but are less critical.
Trade-offs
Building vectors takes effort and needs updating when the model changes.
Pitfalls
Vectors made only of typical rows, which pass while edge cases diverge.