Agentic divergence describes work that moves farther from the Operator’s directed outcome or the evidenced state of the system over successive actions. All agentic models produce generative output at the turn level, and no evidence indicates that the rate of false-verdicting or error-verdicts is zero. A nonzero rate is therefore assumed. Differentiation between volatility and non-volatility of agentic divergence is pending further evaluation.
During a repair, an agent can build an increasingly coherent account of what broke while following a branch unrelated to the actual fault. The explanation may improve internally as the actions it drives move the repair farther from the requested outcome.