Twelve months in, we've shipped, broken, and re-shipped enough features to know what's worth fighting for. These five rules are non-negotiables.
1. The AI never recommends a dose.
Every system prompt has this constraint baked in. Doses live with the clinician, full stop. The cost of being wrong is too high, and the benefit of being right is too low — clinicians can look up doses faster than they can vet an AI-generated one.
2. Every clinical citation must be real, current, and clickable.
Curated knowledge base + retrieval-first generation. Citations are dead links until proven otherwise.
3. Crisis-line numbers always show up for crisis triggers.
PHQ-9 score in the severe band, mention of self-harm, anything that touches suicide — the system surfaces the country-correct crisis line before completing any other reasoning.
4. The clinician can always override, and we log nothing for judgement.
If the AI suggests an MRI and the clinician orders a CT instead, that's their call. We log the override for audit, never for review.
5. The AI's response is always summarisable in one sentence.
If we can't write one-sentence summary of what the AI just said and why, the response isn't ready to be shown to a clinician. It's a forcing function — if the response is too vague to summarise, the prompt or retrieval is broken.