Here are some results after importing the resulting CSV files into a Numbers spreadsheet, and producing a chart:

The jumps in the green line are due to the content of the watch display at the moment the reading was taken; some digit combinations simply scroll by quicker than others. A blank display draws less current, and the voltage reading will be slightly higher. It’s surprisingly sensitive…in another test, the resting battery voltage was seen rising slightly right when the thermostat kicks on in the morning. (Batteries are mildly sensitive to temperature. This is why your smoke alarm waits until you’re sleeping to start that “low battery” beep…the cooler nighttime air brings the already weak battery’s voltage just below the warning threshold.)

Notice at 140 minutes the two lines merge. This is what we dubbed the “death spiral.” The battery voltage has dropped low enough that the watch executes a brown-out reset. This restarts the code, which restarts the display, causing another brown-out…repeating until the battery gives out fully.

(For the record, battery life has been improved substantially since this early graph was taken, so you should get well more than 140 viewings from your watch!)

Here you can see the battery voltage gradually recover following an initial time display. This is chemistry, and is why there isn’t a linear relationship between current draw and battery longevity.

The stuttering parts of the line are due to the limited resolution of the Arduino’s analog-to-digital converter. Those could be filtered out by taking multiple voltage readings and averaging the results, but it wasn’t necessary to go to that level of detail here; seeing the trend was sufficient.