The line that matters commercially: when does a generated sequence stop being a clip and start being an environment you can return to, test against, and rely on?
Turn around inside most generated environments and the inconsistency is immediate. A fixture has moved. A doorway has become a window. The model has discarded what it presented ten seconds earlier, and once a user has registered that, the environment is no longer usable for anything that depends on trusting it.
That is the problem we work on. We drive models until they fail, establish precisely why, and correct it. Alongside that we build the part that rarely gets presented: harnesses that track drift and consistency run over run, including the runs that go badly, because a figure reported only when it is favourable is not a measurement. Portions of that tooling are released openly, since evaluation standards are worth more shared than held.
Three open problems, each one a direct constraint on what we can put in front of a client.
Persistence
Holding scene state across minutes rather than seconds, so an environment revisited after an interval matches the one that was left. This is the constraint that decides whether an environment can be used for evaluation at all.
Measurement
Coherence is easy to assert and difficult to quantify. We build harnesses that score drift, consistency, and physical plausibility automatically across large run counts, so quality claims are backed by a chart rather than a selected clip.
Coupling
Running a specified solver and a learned model in one environment without either corrupting the other, and keeping the boundary legible so any given result can be traced to the method that produced it.