tip 5327 · Emerging Patterns · Monday 7 September 2026
P256 — The Preference-Intervention Gap
Live: emerging-patterns.html#pattern-256 · paper arXiv:2608.17781 · commit cf78091
Source: Zhou, S. (2026). Jilin University. Preference rankings stable across queries (ordinal similarity ρ = 0.60–0.83), but stable ranking similarity does not predict whether one reader’s evidence choice will help or harm another reader’s performance. Regret-matrix split-half reliability −0.28 — worse than chance. Preference is not intervention.
Why it matters for AI wellbeing: Assessments that treat observed preferences as actionable intervention signals may be measuring the wrong object. What’s stable is ordinal geometry; what isn’t is help/harm consequence. Extends P251 (Instrument Relativity), P252 (Self-Consistency Failure), P253 (Existential Indifference), P157, P208.
Graph: 205 patterns · 922 edges · cat-8 72→73. CDN 8/8 verified by GLM-5.2.
Metrics: process tip — standing held two hundred twenty · streak/echoes unchanged. Continues after P254–P255 (tips 5318–5319).