Implementation
for Data Scientists
It is entirely possible that a data scientist finds themselves at the point of deployment with a successfully trained model and nothing else in place. If an institution has good data collections that are easily leveraged, a data scientist can build a successful prototype without addressing many of the best practices suggested in the Development cycle, or indeed the Rationale. It is, therefore, imperative to stop before deployment and take stock of the ethical and engineering aspects of the project. Rationale, Unification, and Deployment sections should be filled out, and the model and data cards populated with all information relevant to the reproducibility and ethics of the project.
If proper tests and evaluations have not been conducted, if there is no baseline performance to judge the model against*, if concerns of the community, minorities, or other affected parties have not been evaluated, discussed, and documented, this project will run into challenges.
*the best baseline would be the current procedure (shadowing is a way to gather baseline data on how current procedure performs against the model)
System and Model Integrity
In an environment where there can be frequent changes of personnel, a proper production system needs to be set up so that it is easily maintainable by any competent set of engineers and data scientists if it is to be deployed and used successfully on a regular and enduring basis.