The claim
U27: “use higher order latents to characterize the more abstract similarity we need to invoke here”. A73: “similarity in the highest-order latents can survive substantial differences in the lower-order latents”, with D_prag = d_Z(Z(u, c_u), Z(v, c_v)), a distributional version, and a dynamical one, D_dyn = E_(a∼π) d_Z(F_a(E_φ(u)), F_τ(a)(E_φ(v))). Its thesis: “Pragmatic fidelity consists in graded preservation of higher-order relational and dynamical structure across successive communicative reconstructions.” And: “The deepest form of communicative drift is not necessarily a change in what is being talked about, but a change in the relationship between the speaker, the audience, and what is being said.”
A73’s identifiability caveat, citing Locatello et al. (2019): there is “no uniquely recoverable decomposition … The solution is to define the latent variables through independently testable communicative contrasts.”
Why it was replaced
U28 (CLAIM-042): “of course, this isn’t a strict hierarchy”. Latents were demoted to “compressed representations of recurring patterns across these observational constraints. They are not the ontological foundation” (A75), which is where the manuscript leaves them (§2; CLAIM-046). The relationship-change line survives in the Abstract (“retain its topic while reversing the speaker’s stance”).