Every part of an expression (each word or role placeholder in a sentence) carries a hidden identifier that keeps its annotations and custom attributes attached to the right part of the sentence.

In some models that identifier is not unique: two or more parts, sometimes in completely different expressions, carry the same one. It happens whenever an expression is copied rather than typed:

• duplicating an expression (Duplicate FTE/OTE, converting a fact type expression into an object type expression) - the new expression got a fresh identifier, but its parts kept the identifiers of the original;
• merging or importing another model - the parts of a newly created expression kept the identifiers they had in the source model, so importing the same material along a second route produces collisions;
• transforming a model (GLR substitution) - copied sentence parts kept the identifier of the part they were copied from;
• loading a model from a SQL repository where the column was never filled - all parts end up with the same empty identifier.

Nothing warns about it, because unlike object-fact types, roles and expressions, the expression parts are not checked for uniqueness. The consequence is that annotations on expression parts cannot be reliably traced back to a single part, and anything keyed on that identifier (documentation, exports, repository synchronisation) may attribute them to the wrong sentence part.