The library problem and the discovery problem
PapersGPT assumes the hard part is already done: you found the papers, they are in Zotero, and now you need to read them faster. Inside that assumption it is excellent, and running a local model over your own PDFs is a genuinely good answer for confidential work.
Lune assumes the opposite. The papers that would change your argument are the ones you have not read, so they are not in the library. Nothing inside Zotero can find them.

