The main issue is we know IQ is meaningfully correlated with real world outcomes, but also that real world environments are meaningfully correlated with ranges

The main issue is we know IQ is meaningfully correlated with real world outcomes, but also that real world environments are meaningfully correlated with ranges of (non-normalised) IQ in a way we can't adequately explain or replicate. It's less obviously correlated with prior knowledge than most tests, but still very trainable (especially if you're a neural network with no generalised reasoning ability whatsoever...) And like all types of performance test the "something" it measures is a vector of pattern matching ability (both the component generalisable to solving more difficult problems and the component that doesn't except in the domain of career-relevant tests) but also other things obviously correlated with future performance but not fitting most people's definition of intelligence (attitude towards test, comfort in guessing, time management in timed tests, and to a degree experience), so whilst we know what the scale of the apparent impact is, we don't know really know what's behind the cause