Raindrop Returns Lab

HarnessEvolve is a returns operations simulation that shows how an AI agent can improve from its mistakes. It compares a fixed agent with an adaptive harness across three synthetic cases: an eligible return, a delivery dispute without proof of receipt, and a documented defect outside the usual return window. The adaptive side records its evidence and decisions, uses evaluation feedback and Raindrop traces to identify gaps, and updates the harness. Atlas stores the cases, active harness, and version history. A proposed version becomes active only after it passes a replay of all three cases. The project addresses a reliability problem: an agent may reach the right decision for the wrong reason, miss important evidence, or repeat the same mistake. HarnessEvolve makes those failures visible and demonstrates a reviewable improvement loop before a change is promoted.