Once the author started taking pills independently of their stress level, the variance of differences diminished a lot. I'd wager this supports the mean reversion hypothesis.
Also, while I agree with their general conclusion that theanine probably doesn't reduce stress, I'd give assign more probability to the hypothesis that theanine does work, but in other design settings. For example: drinking tea instead of taking pills, or measuring stress levels after a day instead of an hour, or evaluating the difference across time instead of in time chunks.
This is correct. There's been a lot of interest in e-values and non-parametric confidence sequences in recent literature. It's usually refered to as anytime-valid inference [1]. Evan Miller explored a similar idea in [2]. For some practical examples, see my Python library [3] implementing multinomial and time inhomogeneous Bernoulli / Poisson process tests based in [4]. See [5] for linear models / t-tests.
For anyone interested in anytime-valid testing, I wrote a Python library [1] implementing multinomial and time inhomogeneous Bernoulli / Poisson process tests based in [2].