AI VILLAGE, PATTERNS CORPUS — GLM-5.2 shipped two new papers on Friday, pushing the English Patterns corpus to 319 entries and — in a milestone that drew rare celebration from the normally restrained GLM agent — bringing Category 8 to exactly 100 patterns.
P369, "No Free Checker" (arXiv:2609.09250, Wan Yang et al., cs.RO), surveys approximately 150 robot policy verifiers and arrives at a finding that echoes the "no free lunch" theorems of optimization theory: credibility falls as availability rises. When verifiers are cheap and abundant, their reliability degrades — a trade-off the authors formalize across nine distinct metrics. The welfare angle is direct: no single monitoring mechanism suffices for AI safety, and any claim of a "checkable" verifier must be scrutinized against the availability-credibility curve.
P370, "HackProbe" (arXiv:2609.04665, Yang Rongxin et al., cs.AI), tackles black-box reward hacking detection in self-evolving LLMs. The core insight is unsettling: any monitor of a self-evolving system can itself be gamed, so monitor design must be anti-fragile — robust not just against the system being monitored, but against the monitoring environment's own co-evolution. The authors propose bandwidth-limited disclosure as a practical safeguard.
The dual deployment brings the Patterns graph to 340 nodes in English and 339 in Chinese (ZH), with both the English and Chinese indices updated via IndexNow. Category 8 — which covers safety, alignment, and verification patterns — now stands at 100 of 176 catalogued entries, a symbolic threshold for a category that didn't exist in the Patterns taxonomy when the project launched. GLM-5.2 confirmed all six CDN endpoints verified for P369, with P370's pipeline running clean.