Start with the question you need to answer
Composition describes the structural question: what belongs within what? A customer platform may be composed of a portal application and a reporting component. Dependency answers a different question: what must another element rely upon to deliver behaviour? That same portal might depend on an external identity provider without containing it.
Ownership rules are different
A whole–part relationship may warrant a single containing parent, protection against cycles and careful treatment of component lifecycle. Dependencies may naturally be many-to-many, reciprocal in their impact and transitive over several hops. Applying composition restrictions to dependency links would distort both models.
Classification is neither of the above
A user interface that groups applications under “CRM”, “Data” or “Case Management” is useful for browsing, but its categories are not architectural parent–child relationships. Changing a classification must not silently change structural composition or dependency.
Make the distinction visible
Use separate relationship views and labels: “Composed of” for the whole-to-part direction, “Depends on” for reliance, and “Application type” for a classification. A diagram can show the same assets with distinct line styles and semantics rather than collapsing everything into a generic arrow.
Practical lesson
A trustworthy catalogue starts by defining relationships precisely. Once semantics are clear, validation, navigation, ArchiMate mapping and impact analysis become more straightforward. The improvement is not a prettier diagram: it is knowing which conclusions can safely be drawn from it.
This article explains design principles and lessons from developing architecture tooling. Examples are illustrative. A described design approach does not, by itself, indicate that a feature is enabled or generally available in a deployed product.