1. Deterministic facts
Chart casting, corpus matching and citation assembly are deterministic. The same input, school and engine version produce the same chart; each chart carries a checksum.
2. Optional polishing
A language model may be used only to rephrase the reviewed L2 text. It must not add facts, new citations or new chart symbols. If the guard rejects its output, the service falls back to the L2 text.
3. Grounded chat
When chat is enabled, the model receives the user's unlocked reading (report text, reading notes, assumptions, citations, gaps and disclaimer) and recent conversation turns. It may explain and connect that material; it is instructed not to cast a new chart, add facts, or provide medical, legal, financial or crisis advice. A visible AI notice is shown in the chat panel, and every reply is stored with the conversation until the user or retention policy deletes it. Messages that indicate a crisis are routed to local support resources before any model call.
4. What the reader sees
If a polished passage is actually used, the report includes a visible AI notice for that locale. If no model was involved, no AI notice is printed. Machine-readable metadata (`llm_used`, `polished`, `notice_version`) is attached to the reading record.
5. What we do not claim
We do not claim scientific validity, predictive accuracy or that a model understands your situation. Readings are entertainment and reflection aids.
6. EU AI Act transparency
Where the EU AI Act applies, the operator must ensure the visible and machine-readable transparency duties are met for the deployed configuration. This template is not legal advice; final wording requires review.