Shownotes
A fixed error in an AI test does not prove the error is gone from the system. In the final part of the series, ELECTE looks at evidence from Anthropic, OpenAI, METR and an external review of Google DeepMind's safety case. Corrections can depend on context. Tests have to look like real work. Monitors need testing too. Why a measured improvement should not automatically mean more autonomy, and why a reliable agent must be able to leave a task unfinished.Send us a text. AI Frontiers is produced...

