Distance from Gray.

The default state of any unstructured or probabilistic output: variable, hard to audit, not production-ready. Gray is not wrong — it is unfinished. Better models emit better gray: the variance narrows, the plausibility improves, and the fundamental property stays.

The output looks right. Nobody can promise it is right. So a human checks it, and the checking costs roughly what the AI saved. The bottleneck in every production AI workflow is not the model — it is the promise.

Generation is commoditizing toward zero. The ability to promise that output is schema-valid, auditable, and safe to wire into a production pipeline is scarce now — and gets more valuable as generation gets cheaper, because cheaper generation means more gray in more workflows.

A claim of quality is an opinion; a delta is a measurement. Instrument the seam and read it: how often the raw input would have failed downstream, how often the transformed output actually does. That gap is the product. Everything else is packaging.

Between the model and everything downstream there is a seam where output either becomes trustworthy or does not. A transformation layer lives there: strict schema on the way out, fails closed, unknown input a hard error. Wrong-but-well-formed is solvable. Wrong-and-shapeless can only be babysat.