Null model
A version of your data with the structure you are claiming removed, and everything else left intact — run through the identical pipeline.
Like you're twelve
You think your lucky shirt wins games. To check, count the wins in the shirt — then count the wins in every other shirt. If the numbers match, the shirt was never the reason.
The comparison is the evidence. A number on its own is not.
Precisely
A null model destroys exactly the structure being claimed and preserves everything else. If the claim is that two series are coupled, rotate each series independently: cross-coupling dies, each series keeps its own autocorrelation. If the claim is about labels, permute the labels. If it is about pairing, shuffle within rows and keep the row geometry.
Run it many times to get a distribution rather than a single comparison, and report your measurement beside it. Where the reported effect is the largest of many candidates, the null must also be on the largest of many, or it is answering a different question.
A null is only meaningful when the instrument could have expressed the answer. Confirm the method recovers a result that is true by construction before accepting "nothing there" as a finding.
Related and equally cheap: the floor of your own metric on structureless input. A self-inclusive separability score reads 0.996 on two piles of noise; the same question asked with a held-out split floors near 0.54.