The people you brief are the people who write the code. Nothing is handed down to a junior team once the contract is signed, and nothing is subcontracted without telling you.
Nearly everything published on world models so far is a demonstration. Fifteen seconds, a single path through the scene, cut immediately before it degrades. The version that carries commercial weight is the one still standing after somebody has worked inside it for an hour, and that is a substantially harder problem than it appears from outside.
That is the bet, and it is also why we run a services practice alongside the research. Client work keeps us building systems that have to survive real conditions, which is the only useful test of whether an environment is ready for anyone to rely on.
Four commitments we hold to on every engagement, including the ones where holding to them costs us the work.
If the work you are describing does not need us, or does not need doing at all, you will hear that in the first conversation rather than after an invoice.
Benchmarks come with the runs that went badly attached. A result presented only when it flatters the model is not a result you can plan against.
Documentation written for engineers who were not present, evaluation tooling included, and your team trained on it. Dependence on us is not part of the commercial model.
Delivery is not the end of the relationship. We remain available for the period where systems actually reveal their problems, which is well after launch.