1. Instead of (only) running eva-query-activity (or a similar substitute) as part of eva-items, run it by itself at random times.
2. Log whether an observation was sampled in this random way or if it was sampled as part of eva-items. Extra field in the data, I guess.
3. When a random query fails, log it anyway with a blank (N/A) value.

I figure the first item can be interesting by itself to many users, especially if you add mood. E.g. say you want to measure an average of your mood to see how a life intervention influences it. Naturally, a reliable average must come from sampling at random times.