Basically what Daniel proposed is, that it would be best practice that for every string that refers to a concept, event, thing, person, unless the editor is certain about item identity, a new wikidata item entity should be created.
I could imagine this as a possible and perhaps elegant solution. My concern is the handling of unknown identity and a workflow towards improved identity recognition, not a discussion string versus item.
Could it be that the types http://wikidata.org/vocabulary/datatype_monotext http://wikidata.org/vocabulary/datatype_multitext become redundant then? Is it possible to simplify the wikidata model by specifying that all language-specific strings are to be represented by an item that keeps the translations together? I find this appealing...
Gregor