JessicaH_TX said:

The pattern that distinguishes a bad batch from an exit is behaviour rather than product.

I disagree that testing history is decisive. It tells you what a supplier did when they were being watched. Continuity of behaviour under stress — a late shipment, a failed test, a complaint — is more predictive than any run of good results.

That is the short version; the long version is somebody else's post.

The figures, for anyone assembling their own picture. The four things to check on any report, in order: does it name your batch, who commissioned it, what method was used, and does it report content as well as purity. A report failing any of those is weak evidence regardless of the percentage on it.

Dr.ObesityLA said:

I disagree that testing history is decisive.

Coming at Dr.ObesityLA’s question from a different direction. Inter-laboratory variation comes from real methodological differences, not sloppiness: which reference standard was used and how it was itself calibrated, the gradient and column, the detection wavelength, and where the integration baseline was drawn. A percentage point or two between competent labs is expected. What is not expected is a large gap, and the usual explanation there is that one report is not of the batch in your hand.

One thing that is still open after tony_orlando’s answer:

How people are distinguishing a supplier having a bad batch from a supplier on the way out?

Closing the loop on my own question.

Went back through my own correspondence with an earlier supplier and the warning signs were all there in the replies, weeks before the quality changed. I was reading the product and not the behaviour.