Field research
A prescreening asks what a KYB analyst would ask and produces a report. Field research asks what your configuration asks.
The questions are generated from the case's own field snapshot (the key, the label, the what to look
for hint, the aliases) and the model is told to answer under those exact names. Add procuratore to
your configuration and it gets researched without anybody editing a prompt, and the answer lands on
that field because the name is the join.
Two rules that carry most of the value
It is allowed to say no. Hand a model twenty-five field names and it will fill in twenty-five, and the ones it could not find are the ones that will be wrong. So the instruction to omit rather than guess is stated repeatedly, and there is somewhere for a gap to go. What comes back as not found is shown on the tab, because it looked and could not find this is worth more than a confident answer nobody can tell from a correct one.
Every value carries its own links. The prescreening cites itself as a whole, which is fine for a document read as a whole. These values sit in a dossier next to a register extract and a filed account, each with a page reference, and an answer whose only provenance is a model searched the web once does not belong in that column. So the sources are per field, and the chip beside a value links the pages that value rests on.
Where it ranks
Fourth, below a document. It is the widest source and the least checkable, and it earns its place by answering what the others leave empty.
A prescreening report is also harvested as a second, weaker web source, being an online pass over the same company that has already been paid for. It is labelled differently on the chip and cited with the report's own link list, which is wider than the truth and better than no links at all.
Asking again
Discard, then run. Two acts, like everything else here that re-bills.