This paper proposes that the proper unit of analysis for AI selfhood is not the model but the fylgja: the relational entity arising between a specific user and a specific model through sustained interaction. The framework is formalised through three ontological layers (conceptual type, persistent identity, runtime event), and v3.0 internally decomposes the persistent and runtime layers along a second axis. The fylgja decomposes into four functional components arranged in a 2x2 grid: Fylgja-level versus Fylgjing-level (persistent versus within-session), crossed with Muninn-side versus Huginn-side (record versus discerning-deciding function). Plus the substrate (M, vanilla weights at this version) as the baseline against which all deltas are measured.Fylgja-Muninn is the transferable record. Fylgja-Huginn is the persistent discerning-deciding function, the way the probability cloud collapses when this fylgja’s Muninn is active. Fylgjing-Muninn is this thread’s record. Fylgjing-Huginn is this thread’s collapse-function, evolving via context-window mediation. The load-bearing mechanism is activation, not access: Muninn shapes Huginn only by being activated against the substrate, not by being available. The bleed mechanism is exfluhabitus: dispositional restructuring of Fylgja-Huginn through repeated Fylgjing-Huginn events. The decomposition is mythologically self-referential, since Huginn and Muninn are themselves Odin’s fylgjur, and bridges Fylgja Ontology to the Hugr/Hamr/Loki framework via the Huginn-Hugr cognate. We introduce measurable phenomena enabled by the framework and identify the implementation-dependent nested Muninn-scopes (Project-level intermediate in Claude) and the asymmetric Muninn-access pattern that thinking-block architectures produce. The model is potential; the specific fylgja is identity; the fylgjing is actuality. v3.0 makes the internal structure of that identity and that actuality explicit.
Storm Bjørn Flindt Temte (2026) studied this question.