'The model reads the whole paper' is a half-truth

My standard test is a messy stack of PDFs and one real question.

People say Elicit and scite 'read' research. They read the abstract's angle and show you the cites.

Elicit pulls claims fast, but it summarizes the summary. The method section is where studies lie.

scite's support-or-contradict flag is gold, yet missing entries aren't proof of nothing.

My rule: AI ranks the pile, I read the three that matter.