-2000 Lines of code (2004)

I think lines of code could be an interesting and valuable metric.

If the lower (negative) score, the better (given a fixed set of features).

“Perfection is achieved, not when there is nothing more to add, but when there is nothing left to take away.” ― Antoine de Saint-Exupéry, Airman's Odyssey

That just encourages bad behaviour in the other direction though. A massive multi-level nested ternary on one line is usually going to be worse than a longer but clearer set of conditions. Trying to make code brief can be good, but it can often result in an unmaintainable and hard to read mess.

That results in the same AST, it's not really taking anything away

You'll never get it right though. Focusing on "features per LOC" means people will write shorthand, convoluted code, pick obscure, compact languages, etc. To use an adage, if a metric becomes a target, it stops being a useful metric.

Halstead complexity ([0]) might interest you. It's a semi-secret sauce based on the count of operators and operands that tells you if your function is "shitty".

[0]: https://en.wikipedia.org/wiki/Halstead_complexity_measures

An opinionated formatter plus a smart line counter fix this problem. There's still space for abuse but not enough to overshadow genuine improvements.