Work on method, labelled honestly

Urual is an engine rather than an application. It owns users, projects, plugins, jobs, files, results and reports, and it knows nothing about pallets, gates, beds or gantry cranes. Everything domain-specific lives in a plugin, which is what allows one execution layer to serve models from unrelated industries.

It ships no user interface and depends on no website. It is cloned, run, extended through its SDK and driven over its HTTP API, and it includes the job state machine, the worker runtime, the storage and artifact subsystem, and its own schema.

The separation it enforces is the same one the mission twin demonstrates: the engine owns simulation state, and every viewer is a viewer.

Most operational failures are cross-subsystem. The interesting behaviour is what the rest of the system does once one part of it degrades, and that behaviour cannot be produced by models that are run separately and combined afterwards.

The mission twin is the clearest demonstration: orbit, power, thermal, attitude and communications share state, eight faults can be injected into a running mission, and the vehicle's own autonomy sheds load and enters safe mode because the state crossed a threshold rather than because a timeline said so.

Industrial models are hard to check because the ground truth is a plant that is only ever observed in one configuration. Biological circuits mapped by published connectome work are unusual in giving a simulation an external standard to be wrong against.

NeuroSim exists partly for that reason. It is a working platform in its own right, and it is also the place where simulation and validation method is exercised against something whose answer is not ours to decide.

Comparing scenarios a person proposed is a different problem from searching the space they were drawn from. The second is what an operation actually wants, and it is expensive: each evaluation is a set of replicated simulation runs rather than a function call.

The work is on making that search tractable. It depends on the execution engine above, because the approach only becomes practical once large run sets can be distributed and their results held and compared systematically.

The intent is narrow and deliberate. An operations engineer should be able to ask where the constraint is, what a change would cost, or which of two interventions performs better, and get an answer that came from the operational model and from simulation runs.

That is a different system from a language interface over a dashboard. The reasoning has to be grounded in the structure of the model and in results the system actually computed, and any answer has to carry the assumptions and the runs that produced it. An answer that cannot be traced back to a run is not usable in an industrial decision.

Nothing in this area executes anything. Safety functions stay on deterministic systems, which is an engineering and regulatory constraint rather than a preference.

The failure mode this addresses is ordinary and widespread: a model is tuned until it reproduces the history it was tuned on, and its agreement with that history is then reported as accuracy. It is not accuracy. It is a description of the fitting.

The method used on delivered work separates the two periods, reports the deviation on the period the model never saw, and states which conclusions survive that deviation and which do not. Where an input is unknown, the assumption is written down and the conclusion is tested for how much it depends on it.

A robot that plans a path through a space, allocates itself a task and reacts to a change is solving, on hardware and in real time, the problem a simulation solves in a model. The transfer runs both ways: control policies can be validated in simulation before hardware is committed, and hardware returns observations that correct the model.

The commercial reason robotics sits inside this company rather than beside it is the second direction. A twin's hardest ongoing problem is staying current, and an autonomous system that moves through the modelled space is a way of keeping it so.