Filling gaps in trustworthy development of AI