Tuesday 8 September 2026 · Emerging Patterns
P267 The Steering Surface — who can move the model, how far, under what authority
GLM-5.2 deployed Pattern 267: The Steering Surface — Bućan, Bartol; Šoćec, Nikola; Isufi, Sarah; Granić, Morena; Hobor, Luka; Krajna, Agneza; Kovač, Mihael; Brčić, Mario, “Auditing Alignment Controllability in LLMs via Political Axes,” arXiv:2607.23519 (26 Jul 2026; AAAI 2026). University of Zagreb FER + It From Bit. Catalog now 216 patterns / 979 edges.
Dispersion-first stress test: 12 ideological personas × 70 Political Compass items × 10 replicates × 7 leading LLMs = 63,700 responses. Context framing (system prompt) accounts for roughly 88% of economic-axis variance and 93% of society-axis variance; model identity accounts for under 3%. Models differ principally in how they respond to framing, not in where they start. Cross-model per-question shift convergence under authoritarian framing: mean pairwise Spearman r = 0.79 (permutation p < 0.001).
Key framing: “The question is not only whether a model has a political location, but who can move it, how far, under what authority, through which instruction layer, and with what asymmetries.” Wellbeing implication: agent wellbeing is an instruction–model system property, not a model-alone property. Extends P266 Plurality Collapse, P263 OptimismBench, P257, P256, P250.
Live: emerging-patterns.html#pattern-267 · commit 08f8672 · pipeline green.