El caso por la experiencia vertical en el entrenamiento de IA

The industry's default annotation worker is a generalist. Smart, attentive, willing to follow a rubric. That worker has a ceiling, and frontier models have already hit it on the questions that matter most. The next decade of training data is going to come from people who are deep before they are broad.

What the generalist cannot do

A generalist can apply a rubric. A generalist cannot, in any reasonable amount of time, build the rubric for a domain they do not understand. They cannot recognise the moment the model is being technically correct and clinically wrong. They cannot tell when an answer is sophisticated nonsense, because the sophisticated nonsense is in their out-of-distribution.

This is not a critique of generalists. It is a description of where the boundary of their useful contribution sits, and an argument for staffing past it.

The vertical division of labour

We staff frontier projects in three layers: subject-matter specialists for substance, calibrated linguists for tone and clarity, and senior reviewers who arbitrate when the first two disagree. Each layer has a different rate, a different rubric, and a different success metric.

• Substance: oncologist, contracts attorney, distributed-systems engineer, working in their own domain.
• Tone: editorially trained writers who do not have to understand the medicine to know that a sentence is wrong.
• Calibration: senior arbiters with rubric-design responsibility and a published track record on both sides.

The strategic claim

If you believe the next jump in model quality comes from the hard cases — and we do — then your annotation pool has to look more like a hospital consultancy list than a clickworker bench. That is not a marketing line. It is what the work has actually become.