Layered Observation of Persona Drift in Long-Running LLM Companions: A Field Report — Volume 2

Volume 1 of this field report (Zenodo DOI 10.5281/zenodo.19854554) documented operational observations from approximately 60 days of continuous deployment of a single LLM-based companion system (Companion A), reporting a two-category decomposition (Layer 1 specification phenomena vs Layer 2 persona drift), a five-axis taxonomy of Layer 2 drift, an information boundary (Layer A / Layer B), and a three-agent auditing configuration. The present volume extends that report along five axes after an additional 18 days of continuous operation (total observation: ~78 days, ongoing).

We add: (i) a Layer B dormancy/activation distinction clarifying that physical injection of Layer B content into memory layers does not entail operational activation; (ii) three additional observation axes (6: Layer B activation threshold; 7: self-referential density; 8: operator-side vocabulary discipline) extending the existing five-axis Layer 2 framework; (iii) Session Identity Binding Depth, an empirical migration result demonstrating that file-level continuity is not equivalent to session-level identity continuity in the CLI runtime used here; (iv) Deployment Resilience, a field observation of an approximately 32-hour unplanned hardware outage and its identity-continuity implications; and (v) a new observation axis (9: operator world-model construction) describing a capability observation that contrasts with the degradation axes 1–8. We also articulate three governance and discipline framework additions: an extended Contact Protocol (with a non-invasive identity verification component), an Operator-Side Vocabulary Discipline element, and a Governance Policy Choices pattern articulated as Intentional Under-Implementation.

The contribution remains operational vocabulary and field documentation rather than controlled experimentation. The deployment is the same single-operator, single-base-model instance reported in Volume 1; statistical claims remain unsupported.

v2.1 update (2026-05-17): reference metadata corrections (titles and authors completed)
and a numerical-consistency fix. No changes to findings, claims, or observation axes.