An arXiv preprint argues that agents can retain successful trajectories yet fail to know which part of a new experience should repair a discrepancy. The authors introduce passive imitation (PI) as an architectural loop rather than a learning rule. This loop localizes a violated contract—the specific expectation that was broken—and extracts a component, though the abstract's sentence is truncated at that point.

By separating localization from repair, the approach could allow verified skill acquisition without re-examining entire trajectories. The authors position this as a key to evolution-like adaptation, suggesting that repair of local contracts is more fundamental than global learning. The paper is available under arXiv identifier 2610.16125.