Five patients. Five different gums.

This simulation is entirely deterministic. No language model is involved in generating any of the text. The five patient personas are authored constructs - each variable (treatment type, iron, zinc, hygiene, weeks, nausea) was set to produce a distinct point in the clinical variable space, and the experience description was written by me to reflect what the literature says about that profile.

The decision to use a deterministic scenario picker rather than a live LLM was made for three reasons: static hosting (this site runs on Vercel with no server-side logic - a live LLM call would require a backend and API costs); explainability (authored text can be reviewed and corrected; LLM text cannot be guaranteed to stay accurate); and scope control (building a reliable deterministic version now is better than an unreliable dynamic version later).

How a live-LLM upgrade would work. The §05 algorithm already produces the patient state as a structured object. That object plus a system prompt anchored to the §04 mechanism descriptions would be sufficient context for a capable language model to generate accurate, personalized experience descriptions on demand. The architecture is scaffolded for this - the authored text is a drop-in placeholder, not a structural dependency.

Limitation: these are not real patients. The personas are grounded in the literature but were not drawn from actual clinical observations or patient interviews. They demonstrate the algorithm's differentiation across the variable space - they are not prevalence-weighted samples of the actual chemo patient population. Actual patient interviews or a clinical survey would be the right next step.