Large language models are remarkably good conversational systems. Give one a prompt describing a patient and it can produce convincing dialogue within seconds. That creates an obvious question: why not simply prompt an LLM to act like a patient?
Because believable dialogue and reliable clinical simulation are not the same engineering problem.
Conversation is only one state
A patient simulation contains interconnected clinical information: history, symptoms, examination findings, investigations, diagnosis, disease progression, management consequences and educational assessment.
If those components are generated independently without appropriate constraints, contradictions emerge. The patient says one thing, the examination implies another, the laboratory results suggest something else. For entertainment that might be tolerable. For education, consistency matters.
Separate responsibilities
SYNTAX uses a modular architecture rather than treating the entire simulation as one undifferentiated conversation, with distinct responsibilities for the patient, reasoning, investigations, consultant review and evidence alongside voice capabilities.
That separation reflects a useful engineering principle: different problems deserve different controls. The system responsible for behaving like the patient does not need to be the same mechanism responsible for assessing the learner.
The patient shouldn't know everything
A generic model often behaves like an omniscient assistant. A patient isn't omniscient. They know what they experienced. They may know previous diagnoses. They may remember some medications. But they don't necessarily know the pathophysiology underlying their symptoms. A patient simulator therefore needs controlled knowledge boundaries.
Clinical truth must survive conversation
The challenge isn't generating the next sentence. It's maintaining the simulated clinical reality while thousands of possible sentences are exchanged — keeping history, findings, investigations and progression coherent throughout.
That is closer to a state-management problem than ordinary conversational generation.
Assessment is another system
Then comes the learner. Did they take an appropriate history? Did they identify the relevant findings? Were investigations justified? Was the differential sensible? Was management safe?
Those are assessment questions, not patient-dialogue questions. SYNTAX provides consultant-style review across history, communication, examination, investigations, differential diagnosis, reasoning, management, safety and professionalism — and the score itself is computed deterministically, never by a language model.
Building simulation rather than conversation
Language models can be extraordinarily useful components of clinical simulation. But the product architecture has to answer a broader question: what must remain true throughout the encounter?
Once you ask that, the problem becomes bigger than prompting. You're no longer merely building something that can talk like a patient. You're building something within which a learner can practise being a clinician.