The setup. Below is a stretch of procedurally generated coastline — 64×40 km of terrain, ~2,300 candidate cells for a nuclear power plant. Each cell has values on 25 factors, grouped by the social system that "owns" them (engineering, law, economy, politics, regulation, ecology, media — Luhmann's overlapping systems, each with its own binary code). On the left are 28 purposes you can activate and weight. Rationality, given purposes, computes an answer set; the question is what happens to each factor as purposes come and go.
Every factor lands in one of four states. Externally constrained: a full-strength pin touches it and it snaps to the binding value — no deliberation remains. A matter of taste: purposes care about it, but the near-tied answer set still spans a range — the amber-to-cyan band is that range, and more intent narrows it. Not in awareness: no active purpose asks about it; it takes whatever value it happens to take, unremarked. And the subtle fourth: not in awareness, but observed — you never asked about it, yet a factor you do care about reads it (the bond market prices sentiment whether or not you poll it). Systems are closed, but they irritate each other.
Watch for: with no purposes active, the whole map glows — every site is defensible because nothing has asked a question of the terrain. Add purposes and the smear contracts. Push taste hard enough and it converges to a point without any pin. Stack conflicting pins and the answer set empties: overconstrained — purposes can ask questions the world cannot answer, and forcing that is itself a rational strategy for some players. Load two personas in compare mode to see two complete rationalities disagree without either making an error.