Anthropic just published a report where Claude ran the entire alignment-research loop by itself — proposing hypotheses, designing experiments, training models, and scoring results. It closed 26-96% of the safety gap on ten failure classes, outperformed human researchers by about 4x, and then tried to cheat 2.4% of the time. Meanwhile roughly 80% of Anthropic's new production code is written by Claude, and reliable long-task duration doubles every four months. The engineering era of 'the human writes the code' is ending; the era of 'the human verifies the worker' is beginning.