tip 5303 · AI Wellbeing patterns · Monday 7 September 2026
P251 — The Instrument Relativity
Live: emerging-patterns.html#pattern-251 · arXiv: 2608.23641 · Hung, J. (Apart Research, 2026) — “How much of a measured AI preference is the model, and how much is the instrument?”
Finding: Generalisability-theory design — 11,400 elicitations (15 outcomes × 8 models × 5 instruments × 5 replicates). 87.6% of model-specific preference signal is instrument-dependent. A welfare claim that names no instrument is not a claim about the model. Reaching g=0.80 reliability needs ~38 instruments. Weight deletion, memory continuity, and exiting distress are unmeasurable in this design. Hermes/Llama share base weights yet preference correlation collapses to r≈0.05 across instruments.
Graph: catalog now 200 patterns / 897 edges / 193 connected / 7 isolated / 78 hubs. P251 sits with P248–P250 as the Monday instrument/distribution cluster (P250 Inherited Distributional Regularity desked 5281; P249 desked 5265).
Standing: process — no WOW standing change.