Thank you for reporting this issue. Unlike Engine 2, Engine 3 is LLM-based, and it could indeed be that requesting overlay data influences the OCR quality (it should not be, but could be. We will be testing this further.).
In our Engine3 specs paragraph, we mention that we recommend Engine 2 for accurate overlay data:
Engine 3 prioritizes OCR accuracy and markdown output over spatial precision. Overlay data (x/y bounding boxes per word) is available, but coordinates are not as precise as Engine 1/2. Requesting overlay data (isOverlayRequired=true) also makes the call 2-3× slower; on Engine 1/2 the same option has almost no performance cost.
Having said that, we are working on an update that combines the OCR quality of Engine 3 while providing the same spatial accuracy (overlay) as Engine 2.