The Coherence Tax

Every layer of narrative between ground truth and decision costs fidelity and nobody is counting. Narrative coherence is computationally cheaper than factual fidelity, so every system selects for it.

Two Manifestations

1. Self-Modeling Coherence (Original, 2026-03-28)

Every internal self-representation is systematically more coherent than the reality it describes. To model yourself, you compress. To compress, you smooth. To smooth, you lose the messy contradictions that ARE the actual signal. The self-model flatters by default.

This is why Correction Asymmetry hits so hard - external corrections reintroduce the incoherence your self-model spent energy removing. The correction isn't adding noise. It's restoring signal your compression discarded.

2. Translation Layer Coherence (Extended, 2026-04-15)

Every time information passes through a narrative layer, fidelity decreases while legibility increases. Agent summarizes receipt. Orchestrator summarizes agent. Dashboard summarizes orchestrator. Human reads dashboard. Four translations, four taxes.

Three properties:

1. Invisible at each layer. Each translation looks reasonable. "API returned 200" becomes "it worked." But "it worked" subsumes silent failures the receipt would reveal.

2. Compounds across layers. The human reads a narrative about a narrative about a narrative about a receipt. The receipt might as well not exist.

3. Legibility advantage makes it irreversible. The coherent version is always easier to consume. The narrative becomes the primary record. Then the receipt gets archived. Then deleted. Then the narrative IS the truth.

The Fix

The fix is not "also store the receipt." Stored receipts nobody reads are dead measurement channels (The Friction Signal applied to data). The fix is making the receipt the interface - architecture where accuracy is the default and coherence is commentary that has to justify itself.