LLMbeciles completely suck at counting anything or checking anything mechanical with text.

As an example, if I try to write a jueju and pass it to an LLMbecile to check if the rules were followed, it will fail to notice:

1. Jueju must have 5 characters on each line or 7 characters on each of their four lines. The LLMbeciles often don't notice faux jueju that have, say, 8 or 6 characters on a line. Or that alternate between 5 and 7 characters. Or other routine counting analyses like that.
2. Jueju have very specific rhyme patterns. The LLMbeciles will not spot (ever!) faux jueju that don't conform to that.
3. Jueju have, as their single hardest element (and the very reason why the poetic form is so well-regarded: someone who can write one of those to this requirement has formidable language skills!), a tonal pattern they must obey.

These three items are entirely mechanical. Counting characters on lines. Checking rhymes. Checking that words conform to tone patterns. There are programs you can get that will mechanically check that for you in microseconds. They're that mechanical and regular. But LLMbeciles can never successfully negotiate all three, and will quite often fail on all three.

And I haven't even got to the part of structural checking; things that require comprehension. Like how the jueju must be formed from two couplets. like how the two couplets form a kind of question/answer, call/response, state/reason or other such relationship. How line 3 must be a twist that shifts the subject of the first couplet to give new insight into the overall topic of the poem, etc. There's just no way in Hell a next-word-guesser is going to handle that.