Skip to content
Content modellingCMMF 1.2.0

Translation

Translation for CMMF 1.2.0. Imported from the canonical CMMF repository.

Why Translation Needs Its Own Mental Model

Section titled “Why Translation Needs Its Own Mental Model”

CMMF is designed as a source format, not as a direct declaration of one CMS implementation.

That means translation is not a mechanical copy step. It is a controlled decision-making process that maps source intent into a target platform.

Why Translation Is Not Fully Deterministic

Section titled “Why Translation Is Not Fully Deterministic”

Two translators can produce different but valid outputs from the same CMMF input.

For example, a contained structure may become:

  • an embedded object in one platform
  • a separate content type in another
  • a reusable block schema in a third

The important thing is not that every target looks the same. The important thing is that the source intent is preserved and translation decisions are explicit.

A translator should preserve:

  • model identity
  • field identity
  • content structure
  • relationship semantics
  • meaningful validation intent
  • important editorial metadata where feasible

A translator may need to decide:

  • whether to inline or split structures
  • how to represent contained ownership
  • how to handle unsupported validation features
  • how to represent settings and taxonomy constructs
  • how much to optimise for target-platform conventions

Without a decisions log, translators become black boxes. That makes debugging, governance, and architecture review much harder.

A good decisions log explains:

  • which choices were fixed by spec
  • which were inferred
  • which were target-platform compromises
  • which were lossy

Future translators will likely benefit from explicit modes such as:

  • strict
  • optimised

strict focuses on preserving source intent as directly as possible. optimised allows more adaptation for target-platform fit, while still documenting those adaptations.