We're starting with keyword-matching here (not GPT-3). For the preregistration, multiple comparisons, and intent to treat columns, our internal testing estimated that Elicit finds the answer about 78% of the time if we have the paper and the paper mentions these phrases (recall).
When Elicit thinks it has found the answer, it's correct more than 90% of the time (precision) since we're using keyword matching. But you should always click in to the cell and double check its work!
Note, these columns only work if we have the full text and make the most sense for randomized controlled trials. You can filter for RCTs and sort by the PDF column.