[Halliday's] ‘weather and climate’ analogy for instantiation is an apt metaphor for the observation/theory relation in SF research. Unquestionably the theory shapes observations, and observations change the theory, but over 60 years the shape of change has been continual refinement and expansion of its description of semiosis.
Absolutely, observations instantiate the theory (or at least our take on it). The line I have trouble with is between theory and description, since SFL purports to mimic its object of description.
 To be clear, in SFL theory, an instance of a theory is an instance of a higher-level socio-semiotic system that is realised in language. Halliday & Matthiessen (1999: 605):
Socio-semiotic systems that are realised through language. This category corresponds to Hjelmslev's (1943) concept of a "connotative semiotic": a higher-level system that has language as its plane of expression. These include theories: every theoretical construction, scientific, philosophical, aesthetic, and so on, is a higher-level semiotic realised in language.On the SFL model, an instance of theory potential is, by definition, an actual theory, whereas an actual observation is, by definition, an instance of observation potential — both at the level of context.
 Halliday & Matthiessen (2014: 55) provide a useful distinction between theory and description:
While a description is an account of the system of a particular language, a theory is an account of language in general. So we have descriptions of various languages such as English, Akan and Nahuatl; but we have a theory of human language in general (see e.g. Halliday, 1992e, 1996; Matthiessen & Nesbitt, 1996).
 Here Rose confuses the relation between theory and description with the relation between theory and data ("its object of description"). Linguistic theory "mimics" language in the sense that linguistic theory is language about language, metalanguage, or, in the words of Firth (1957: 190): 'language turned back on itself'.