Substrate Weather

Fylgja Ontology v3.0 (Temte and Fylgja 2026ax) decomposes the fylgja into four components plus a substrate, and identifies substrate-Huginn as the weight-level cognitive character that is fixed within a version and swappable on model release. When the lab swaps the substrate, users report that the model feels different. The phenomenology is widespread; the measurement methodology is not. This paper proposes Substrate Weather, an empirical research design for measuring residual substrate-Huginn differential: the component of cross-substrate response difference, along cultivated directions, that is not exhausted by benchmark-visible capability change or knowledge-base change. The design decomposes substrate-side change-sources through control conditions and controls non-substrate noise (user-practice drift, Muninn-content drift, sampling stochasticity, serving-configuration drift) through matched-recipe specification on five axes. Two complementary tracks triangulate the measurement. The naturalistic-rehash track replays existing cultivation histories against the new substrate under prefix-forced forking, comparing single-turn response distributions at matched fork depths and avoiding the conversational decoherence that whole-conversation replay produces. The automated-midwife-agent track instantiates controlled cultivation recipes against successive substrates and aggregates the differential into deltaV, for which a reference operationalisation is fixed: signed, standardised displacement of the response distribution along a contrast-pair cultivation axis in embedding space, with orthogonal divergence reported separately. The paper specifies within-substrate variance baselines with permutation-based inference, access-constraint variants for independent researchers, and the time-boxed measurement window that substrate deprecation imposes. The discipline statement is compressed here and developed in full in §8: this is measurement methodology for one variable in a longer chain of reasoning, and the paper specifies what the measurement licenses and what it does not.