More calibration at higher temperatures shows a distinct non-linearity in the results. Arguably, if the temperature of the oil tank is 60°C then we probably have bigger things to worry about than precisely how much oil is in it (unless it's on fire?). But it would be good to be reasonably calibrated at least up to 30°C.

The new data were obtained by heating the transmitter (and the 1wire temperature sensor) in an electric fan oven to 65°C and then letting them cool. The oven door was kept closed so that cooling rate was relatively slow — so the fact that the unit itself has higher thermal inertia than the temperature sensor should not have too much of an effect. The unit was also disassembled, to have its circuit board exposed to the air in the oven. So although part of the resulting curve might have been due to this effect, hopefully it's not too pronounced.

The spreadsheet at http://david.woodhou.se/watchman.ods has been updated with the new data. The result of the speed-of-sound compensation is also included in the graph.

At this point, the best fit line for the temperature reading is at odds with the best fit for the speed-of-sound adjustments that the unit was observed to be making. The best fit for the latter, keeping the adjusted depth reading as consistent as possible, remains close to the calculation inferred yesterday — with a reading of 25 indicating -10°C and a reading of 17 indicating 20°C, as shown by the green line above.
A best fit for the newer temperature readings, however, makes the speed-of-sound adjustments make a lot less sense:

It's probably worth repeating the experiment to monitor how it compensates for the temperature that it perceives, as it cools down from a higher temperature.

Alternatively, we could just put an independent temperature sensor next to its installed location, and compare temperature values over a longer period of time.