In the cohesion model your post advocates, texts are modelled instantially, as cohesive relations between grammatical instances. For empirical research, this model must resort to classifying text types by statistical counting of grammatical types (not a ‘bag-of-words', but a countable collection of clause types). It must do so because it no longer has access to the SFL empiricism of systems, since cohesive relations are defined as 'non-structural’, i.e. outside the feature/structure relations of grammatical systems. …
 To be clear, the cohesion model referred to is the SFL model, first articulated by Halliday & Hasan (1976) and integrated explicitly within the lexicogrammatical model in the four editions of An Introduction To Functional Grammar (Halliday ± Matthiessen).
 This misunderstands instantiation. In SFL theory, language-as-text is construed as an instance of language-as-system, and in this sense, all texts are "modelled instantially" by those who understand the theory.
 This misunderstands instantiation. In SFL theory, every deployment of cohesion in a text is a lexicogrammatical instance of the lexicogrammatical system of cohesion. The wording "cohesive relations between grammatical instances" is nonsensical.
 To be clear, empirical research is a way of gaining knowledge by means of direct and indirect observation or experience.
 To be clear, grammatical cohesion and the classification of text types are distinct domains of SFL theory.
 To be clear, in SFL theory, the different frequencies of grammatical choices in texts are instances of the different probabilities that distinguish different registers.
 On the one hand, this is invalid reasoning ("it must do so…") from the series of misunderstandings that precede it; see  to  above. On the other hand, it invalidly argues a false conclusion from a misunderstanding of a true premiss. The argument is as follows:
- A. cohesive relations are defined as non-structural (true)
- B. cohesive relations are outside the feature/structure relations of grammatical systems (misunderstanding of A)
- C. the model of cohesion has no access to systems (false)
To be clear, the reason that B is a misunderstanding of A is that the fact that the syntagmatic realisation of cohesion is non-structural does not entail that the paradigmatic organisation of cohesion is not modelled as system. The claims of both B and C are falsified by the system of cohesive conjunction (Halliday & Matthiessen 2014: 612):