Suggestions for tools/resources for producing better MathML

People sometimes send me an expression that doesn't speak well with MathCAT
or even occasionally, causes a crash. Invariably, the MathML is technically
correct but semantic nonsense with variables or numbers randomly being in
mtext, sometimes runs of them into a single leaf node, etc.

One person asked me if there was a simpler document than the spec that they
could point their vendor (the company generating the documents with bad
MathML) on what is "good" MathML. Does anyone have some suggestions?

    Neil