the subjective framework is exactly why its good

prior bms relied mostly on unit tests or synthetic judges which are easily benchmaxxed, which leads to nobody trusting benchmarks

we need people manually checking the data for good code quality