SkillWeave: Evidence-Driven Co-Evolution of Instructions, Code, and Workflows for Scientific PDF Translation

Agent skills in production systems are jointly determined by natural-language instructions, executable code, persistent workflow state, and verification logic. Scientific PDF translation exposes the interaction among these components particularly clearly: a fluent target paragraph can coexist with omitted source prose, damaged mathematical notation, incorrect object ownership, or a misleading completion status. SkillWeave turns such failures into evidence-driven updates at the appropriate system layer. The framework traces each observed defect through source extraction, translation catalogs, accepted text, layout composition, and rendered artifacts; it then records whether the resulting change belongs in instructions, code, workflow control, verification, or a document-local repair. The evaluation covers four engine versions, six fixed fixtures, 48 mechanism probes, seven full or ablated conditions, eight native-region restoration trials, and document-level case studies.