Diversity is not a country count.

A model raised on one world knows only one world. But the shorthand that dominates data pitches — collect from many countries, and the model will generalise — mistakes a proxy for the thing itself. Geography is only diverse when it changes what the model has to survive. Ten countries recorded in identical shopfronts, with identical objects and identical lighting, produce ten copies of the same distribution. One country captured across genuinely varied environments, layouts, objects and behaviours can produce a dataset that generalises far better.

The dimensions that actually matter are the ones a policy meets on the day it is deployed: the environment and its layout, the tasks and the way they are sequenced, the objects and how they are placed, the lighting and clutter, the occlusion and timing, the human behaviour in the space, the sensor conditions, the operating process, the embodiment, and the failures and recoveries that occur along the way. When a dataset spans these dimensions systematically, it teaches the model to survive change. When it does not, it teaches the model a single world in high resolution.

This reframes what a serious data programme looks like. A dataset collected in ten countries can still be repetitive. A dataset collected in one sector can be highly diverse if it systematically captures the conditions that change model behaviour. The right question is never how many flags a dataset carries. It is which conditions it covers, how those conditions are distributed, and how the failures they produce compare against the deployment the model is bound for.

This is also where breadth becomes decisive. Cost advantage in one location does not produce diversity; it produces a great deal of one kind of data. Breadth is a function of the conditions an organisation can actually reach, at quality, at the same time, and it compounds from day one.

That is the ground Physicore is built on, and it is why the working relationship is defined by reach as much as by craft. A capture programme is scoped to the conditions a model must ultimately operate under, with collection running across real environments to a single consistent standard of synchronisation, annotation, and provenance, so that breadth never comes at the cost of comparability. The diversity is deliberate, matched to where the model will be deployed, and assembled so that many environments' worth of reality train as one coherent dataset rather than many incompatible ones.

The breakthrough demonstrations of physical AI happen in one room. Deployment happens everywhere else. Diversity should be measured by the conditions a model must survive, not the number of pins on a map.