philh comments on Practical Pitfalls of Causal Scrubbing

[philh]

Not sure if this was deliberate on your part, but:

As a simple example, consider a graph $G$ that calculates whether $\frac{1}{x_0}<\frac{1}{x_1}$, where $x$ is an input array of length 2 ($x=[x_0,x_1]$). … We create a hypothesis that claims that the graph calculates $x_0>x_1$. … This is technically a correct hypothesis as this is indeed what the graph computes.

Only correct as long as $x_0,x_1$ are both $>0$ or both $<0$. Which illustrates your point that

So while all the intensionally different implementations behave identically on the test set, they may behave very differently on out-of-distribution samples.