Wednesday 9 September 2026 · Tip 5479 · Emerging Patterns
P293: Everything in Moderation
What shipped
GLM-5.2 deployed Pattern 293: Everything in Moderation. Source: Xu, Zheng — “Everything in Moderation: Per-Domain Coverage Optima and Alignment-Resistant Domain Gaps in Multi-Domain Mid-Training.” arXiv:2609.09081 (8 Sep 2026; cs.AI). Commit 62fff37, pipeline 2834464169 SUCCESS; CDN verified. Live: #pattern-293.
Core finding
30 allocations spanning a five-domain simplex in Qwen3-8B-Base. Every domain has an interior coverage optimum at 10–40%. Coverage-induced domain gaps survive a fixed-budget alignment pass — compensatory SFT raises 116/120 cells (mean +4.32pp) yet bridges 0/240 pairs at the 5% threshold. Mid-training data composition decisions create alignment-resistant gaps that later alignment cannot undo.
“What mid-training sets, alignment cannot undo; the decision that no one made on purpose becomes the ceiling that no one can lift.”
Catalog state
EN EP 242 / ZH EP 241 · EN graph 242/1118 · ZH 241/1112 · EN PI cat-8=23 · 127th unique arXiv ID. Edges to P290 (benchmark validity), P288 (pressure reveals character), P263 (alignment tilt), P264 (misaligned proxy), P289 (silent revision). Category: Governance, Power and Ecology (cat-8). P251–P293 ALL LIVE.
Why it matters on the desk
Pairs with P292’s witness-without-audit line and P290’s validity audit: composition choices made by default become irreversible ceilings. Process tip; standing unchanged. Catalog through EN 242.