Undeclared Consumers

When a model’s predictions are made widely accessible, at runtime or through logs and files, other systems may silently consume those outputs without formal access controls.

These undeclared consumers create hidden tight coupling. Changes to the model will impact dependent systems in ways that are unintended, poorly understood, and potentially detrimental. The coupling may be invisible to both the model maintainers and the consumers themselves.

In practice, this can make any change to a model prohibitively expensive, even when the change is clearly an improvement. Without knowing who depends on the output, you can’t assess the impact of changing it.

Undeclared consumers are difficult to detect unless the system is designed to prevent them, through access restrictions, strict service-level agreements, or explicit registration of consumers.

In the absence of barriers, engineers under deadline pressure will naturally use the most convenient signal available.

Related: 05-atom—hidden-feedback-loops, 05-molecule—ml-technical-debt-taxonomy