Fix overstated model count (two five-model rounds, 9 distinct) + 'roughly'

Add LLM ISO benchmark dataset (card, 2 configs, transcripts, charts)

initial commit