FINSKILLOPS reframes agent self-improvement as scoped skill patches that must earn deployment

arXiv·low signalXBlueskyLinkedInCopy link

arXiv 2609.19680 (submitted 2026-09-17) targets SEC filing QA, where new questions repeatedly expose heterogeneous errors in period, entity, evidence use and calculation, and existing self-improvement methods give little control over where a correction applies or which previously correct answers it breaks. FINSKILLOPS derives reusable skills from evidence-grounded typed failure diagnoses and governs them through targeted validation, treating post-deployment improvement as controlled behavioral maintenance rather than continuous learning. It arrives the same week as SkillAA with the same core move, making regression scoping the gate on any skill edit.