At breakfast I went off on my complaint that if you have humans integrated into the operational decision-making of any astronomical project, you can destroy the statistical or legacy value of your data. This is a complaint I have had about some extreme-precision radial-velocity projects: Some investigators make decisions, before or during an observing run, to maximize planet yield that involve people in a room, talking. And they are talking about the outcomes of previous observations. That leads to sample selections that depend on the data in complex and un-model-able ways. As in: You would need a complete simulation of the human investigators and their group decision-making even to simulate the observing process, let alone build a probabilistic model of it. That's been a disaster for radial-velocity experiments, and one of the reasons that most of the populations inferences for exoplanets have come from Kepler. Of course, in their defense, many of the radial-velocity projects were trying to maximize planet yield, with statistics be damned!
This made me realize: We should look at things in the experimental-design and active-learning fields to see if we could obtain all the planet-yield benefits that exoplanet projects get from adaptively changing their observing mid-stream, but retain the usefulness of the data for populations inferences, by going to simple algorithmic replacements for the proverbial smoke-filled room. My intuition is a big yes. Another intuition is that any adaptive observing algorithm will necessarily be tuned to a science goal, so you will have to decide in detail what you value in addition to planet yield.
At Stars Group Meeting, John Brewer (Yale) described in detail the EXPRES radial-velocity spectrograph that is being built and commissioned for the Discovery Channel Telescope. He described design decisions, with many valuable contextual comments. One of these is that gas-cell spectrograph designs seem to be declining in popularity, while rigidly controlled lab-bench designs (so calibration parameters vary little and slowly) are rising in popularity. There are many radial-velocity projects getting started right now, so the landscape for this kind of work is changing fast. (Hence the above comments about experimental design!)