Do benchmarks even represent the real life environment workflow?

Don't get me wrong, I am aware that latest benchmarks do represent the actual workflow much better than the previous benchmarks. But, when I was seeing the methodology of evaluations in Artificial Analysis, I noticed something - In many evaluations, models are highly restricted - no internet usage, no tools usage, limits on number of steps, time limits, etc. Subjective grading - Irdk what to say about this... How are they even using LLMs to judge? (Are LLMs even capable of judging?) Then we have evaluations where humans judge the quality. I mean, we all have different tastes, different quality requirements .. how are they even judging?? Custom harness - We all know how much harness do matter. But these evaluators have their own custom harness (like stirrup) which no one even uses in their workflow. Public dataset - There are many benchmarks whose data is public and yk trainers often try to benchmaxx. The best evaluation benchmark will be where the dataset is private and there is objectified defined grading and models have no unnecessary restrictions (like limit on internet usage, tools, steps, time, etc) along with their own harness (or most widely used harness systems). submitted