AI-assisted development has broken agile software engineering’s core assumptions. Developers report feeling 20% more productive while measured performance declines by 19% (METR 2025). Organizations see individual output rise while delivery velocity shows no improvement (Faros AI 2025). Current theory lacks the vocabulary to diagnose why. We model software behavior along two axes—specification and verification—yielding four categories: Sv, Su, Ev, Eu. The key insight: behaviors without deliberate human decision are not specification, regardless of documentation or testing. AI generates implementation, tests, and specification-like documents at machine speed—but none is specification until a human decides it should be. Four contributions: (1) a two-axis model with Ev (verification without specification) as novel concept; (2) the verification paradox; (3) value-verification as a distinct axis; (4) structural analysis showing that when implementation cost approaches zero, the management challenge shifts to specification and verification.
Franny Philos Sophia (Mon,) studied this question.