Preprint describes single-call method for root-cause consensus

A preprint describes Loom, a framework for combining conflicting free-text diagnostic hypotheses in root-cause analysis. The system projects hypotheses into an embedding space, iteratively reweights them around a centroid, then uses one lightweight LLM synthesis call. On the OpenRCA benchmark, the authors report that Loom matched an autonomous agent on two datasets and trailed it on two others while using one LLM call per incident and running about 26 times faster. The paper also reports deployment lessons and negative results in redundancy detection.