My checklist whenever a new v3 submission domes in: * What model(s) is it built off of? * How targeted is it? * What's its training data? * How much does it cost/take to run? * How’s memory handled? * In their words, in 2 sentences, what is the secret sauce that makes it work?