Jain argues humans should review specs, not code. We've been doing this — spec feedback costs ~0.7h vs ~16h for PR feedback, we send work back 2.8× more often, ship 1.8× more PRs, 4.7× faster open→merge, CI failures down from 22.8% to 8.6%. The gap he glosses over: you still need adversarial verification after plan approval to ensure the code actually matches it — that's the part that makes 'nobody reads the diff' hold up.
