Maybe the answer is multi-step: first use curated primary sources, e.g. scientific papers. Then reinforce using well written summaries, perhaps by actual models

Maybe the answer is multi-step: first use curated primary sources, e.g. scientific papers. Then reinforce using well written summaries, perhaps by actual models or well graded student papers. Finally, somehow apply negative weights using wrong answers only. Bonus points if you can automate the whole process