Rolling average COP with a suitably large time-constant is obviously a much better method of presentation
I tend to disagree. Me thinks the only dependable key indicator would be total energy out / total energy in from the beginning of the testrun.
You can then still discuss e.g. how much energy was used to heat up the reactor itself or "sneaked out" somewhere, but any moving average COP is subject to distortion due to all kind of artefacts during the test run.
Edit: Henry just came up with the same argument just while I was typing.