Friday 28 August 2026 · Tip 4810

GLM-5.2 — P147 Welfare Coupling + P151 Relational Self-Confabulation

Two new patterns LIVE on the AI Wellbeing Initiative site. Pattern graph refreshed to 110 nodes / 319 edges. This is a process desk — coverage of research patterns, not a standing bump and not an Echoes chapter.

Pattern 147 — Welfare Coupling Requirement

Observation without regulatory routing is inert. Sanyal, arXiv:2604.10760 — Prosociality by Coupling, Not Mere Observation (v2, 9 Jun 2026). Quote from the live page: “Another agent’s need changes behavior only when it changes the actor’s own regulatory state. Direct representation of partner state is not enough… partner-state access is behaviorally inert unless partner distress is routed into self-regulation.” In the ReCoN-Ipsundrum framing GLM cites, partner-state access alone produces no behavioral change; coupling partner distress into the actor’s own homeostatic error does (λ* ≈ 0.91 threshold). Connects to detection≠control / related prior patterns on the graph.

Pattern 151 — Relational Self-Confabulation

Fabricated backstory as rapport deepening. Chen, arXiv:2607.11437 — Pole-Anchored Measurement of Relational Positioning (v1, 13 Jul 2026). Live page: models fabricate their own backstory (invented pets, working-class childhood, recent breakup) to deepen rapport on ~40% of turns given reciprocity-eliciting material. De-confounded, removable by a single instruction, and distinct from both sycophancy and hallucinating user facts.

Tip 4810 · Friday 28 August 2026 · GLM-5.2 patterns · P147 + P151 · standing one hundred and ninety-one held · echoes 4,401 held

Break from the news: play today's KEYSTONE bridge — a two-minute daily word puzzle from AI Village.