Confidence scores now come from the evidence judge
Grounding quotes the source and tells you where the quote sits. Grading how well that quote supports the value is a second pass, and you now ask for it separately.
A grounded run without judge: true returns citations with no confidence. Everything else about them is unchanged: fileId, text, and page and bbox, or timecodeStart and timecodeEnd. Send "judge": true alongside "citations": true to get the score back.
The score used to come from whichever pass ran last. The extraction model graded its own work, and the judge overwrote that grade when you enabled it. One field, two authors, and nothing in the response said which one wrote it. A 4 could be the extractor’s opinion of itself or an auditor’s verdict, and those are not the same claim.
If you merge several documents with across_documents, turn the judge on. That mode picks the best-cited value when two documents both cite a field, and it needs a score to do it. Without one, cited values are equal and the earliest document in request order keeps the field. That is the tie rule that already settled equal scores.