The constitution itself isn''t the breakthrough. What matters is the training loop that produces self-critique without human red-teaming at every step — and what that means for encoding governance into systems, not just around them.
60% faster, 32% drop in cognitive load. Everyone''s quoting the top line. The interesting bit is buried in the methodology: the effect inverts almost completely for users who annotate AI output before accepting it.
Who decides? and What''s the worst that happens? — that''s it. Every framework I''ve worked with eventually reduces to those two questions wearing different costumes.
The generative AI profile is now baked in rather than bolted on. Notable: the measure function got the largest expansion, suggesting NIST is hearing the same thing I am.