HOST: So if I'm using AI assistance for documents and charts, why should I care?
EXPERT: You might eventually want one assistant's useful notes to help another tackle a similar question. The authors tested that idea on set questions, not at a workplace.
HOST: So what kind of note could move between them?
EXPERT: For example, a note might say to check a junction on a map and then point to the image detail behind that advice. That's an illustration, not a reported result.
HOST: Does the assistant have to learn a new answering model in this case?
EXPERT: No, Epicon stores lessons outside it. One small model revises notes, another organizes them so a later assistant can find them.
HOST: So what actually changed in the transfer test?
EXPERT: The bank came from Codex with QUEN 3.8 27B. The receiving system changed either the answering model, called the backbone, to Gemma 4 31B, or the surrounding tools and coordination software, called the harness, to DeepSeek Harness.
HOST: So, looking at what the authors themselves report, what did they actually observe, and what do you think we should avoid assuming beyond that?
EXPERT: Average benchmark scores rose in both transfers with one attempt per question. Some individual tasks still declined. The construction and test questions were separate project splits from the same benchmarks, and this preprint does not show a workplace deployment.