EvoPilot preprint reports verified gains from agent-led video-search experiments

Researchers report that EvoPilot, a human-gated system for agent-led model development, found a 3.20-percentage-point offline retrieval improvement during a 37-day campaign on an index of hundreds of millions of videos. A separate seven-day randomized evaluation estimated a 0.66% relative increase in a retention-oriented search-quality measure. The method assigns specialized agents to experiments while deterministic checks and durable records reject invalid comparisons. Those controls exposed a pre-existing evaluation defect that an earlier automated attempt had wrongly blamed for a 22-percentage-point decline. The results are presented in a preprint and concern one video-discovery system.