Listing the sources and why they were included before any synthesis is a strong constraint, and it matches my experience that report trust drops as the source count climbs past a few dozen. Your note that Flash tier models resist pulling the original graphic or quote while Pro tier models comply suggests the instruction fights the habit of paraphrasing. Have you tried a cheap verification pass that checks each pulled quote actually appears in the cited source?
Look forward to trying it out! Also a very common use case for me and I’ve definitely gritted my teeth in frustration more than a few times (and wondered if it’d be more efficient to search for sources myself) at the pitfalls you mention.
Listing the sources and why they were included before any synthesis is a strong constraint, and it matches my experience that report trust drops as the source count climbs past a few dozen. Your note that Flash tier models resist pulling the original graphic or quote while Pro tier models comply suggests the instruction fights the habit of paraphrasing. Have you tried a cheap verification pass that checks each pulled quote actually appears in the cited source?
Haven't found it necessary as I end up reading most of the sources as well. I also haven't found inaccuracies while doing so.
Look forward to trying it out! Also a very common use case for me and I’ve definitely gritted my teeth in frustration more than a few times (and wondered if it’d be more efficient to search for sources myself) at the pitfalls you mention.
Very interesting from a HS educational perspective. TY