Can a decision model triage email better than an LLM? Clef vs Claude Haiku on a DGX Spark
Every email that reaches a LibraOS desk gets triaged: who sent it, what it is, what the reader has to do, which part of the business it belongs to, and whether it looks risky. Today that runs on an LLM. We tested Cloudflare's new Clef decision models against it: Clef 27B and the smaller clef-flash, running locally on a single NVIDIA DGX Spark (GB10). The short version: Clef 27B found more of the mail that needs a person, while the LLM kept more noise out of the inbox. For now the LLM stays in charge.
