"If you're using AI, how do I know it's not making things up?"
Fair question. Here's the honest answer.
We use large language models for the natural language command interface — your bridge crew speaks to you in plain language, interprets your intent, reports tactical situations. That layer is LLM-powered.
The universe is not.
Underneath the conversation sits a fully deterministic physics simulation. Newtonian gravity. Conservation of momentum. Thermodynamics. When your Navigator plots a course to Mars, the fuel cost isn't generated by a language model. It's calculated from your ship's mass, drive efficiency, and orbital mechanics. The math is the math.
These two layers are architecturally separated. The simulation engine has no knowledge of natural language. The language model has no authority over physics. Your crew can discuss, recommend, report, and occasionally be wrong about their interpretation of sensor data — just as a human crew would. What they cannot do is change the underlying reality. They can't invent thrust. They can't hallucinate fuel into your tanks. They can't make your ship move faster than the physics allow.
How we mitigate AI errors
Are we still solving problems? Yes. During alpha testing, we're discovering the specific ways language models introduce error at the narrative layer, and we're building constraints to narrow that surface. A crew member might misread a tactical situation. They won't misread the laws of motion, because they don't have access to the laws of motion. The simulation owns that.
This distinction matters because it's the same distinction that makes every serious engineering system trustworthy: separate the control plane from the data plane. Separate interpretation from truth. Let humans (and language models) handle ambiguity, judgment, and communication. Let deterministic systems handle physics, economics, and state.
The crew is the interface. The universe is the authority. These are not the same system, and that's the point.
← Back to Comms