I know what you mean, but it's also easier to fake. I've definitely run into the situation where someone's projects look good, but then they are terrible on the job. I'm left to conclude they were pairing with someone more knowledgeable or cloning some idea or maybe has spent years preparing to tackle that one problem and are not good in general problem solving.

Not saying you won't get a lot of false negatives with high-pressure on-site coding or whiteboarding, but if done correctly I think it's probably the most signal you can get for the least time. Granted it requires good judgement from the interviewer, and a lot of candidates to calibrate. It's also unfair to candidates who are easily rattled, and biased towards candidates who have time to do a lot of preparation. In the past I have preferred contract-to-hire as the gold standard if that works for a candidate, but that only works if the pipeline is small.