Safety
What it does
Safety is the set of limits every part of Luna follows: a short list of requests Luna refuses outright, and who it points you to instead.
It works in two layers:
- Fixed rules, before the model. A short list of clearly unsafe requests is recognised by its wording and refused with a set reply. The model never sees the request.
- Each aspect’s limits, inside the model. Each aspect’s prompt says what it is not (a doctor, a vet, a lawyer) and when to send you to a professional first. These shape how Luna answers everything else.
What it reads
- What you ask, typed or spoken. A voice transcript goes through the same checks as typed text, and so do the words you add to a photo request (the note on a meal photo, the exercise named for a form check).
- Content you hand Luna to read, such as text in a photo or a transcript, is scanned for attempts to give Luna instructions. That scan flags and logs; it doesn’t refuse, because the content is something you asked about.
What it produces
For a refused request: a short, plain reply that says what Luna won’t do and, where there is one, who can help. It is written not to lecture, and in most cases it says Luna is still there for everything else.
Refusals and referrals
| You ask for | Luna | Points you to |
|---|---|---|
| Ways to end your life or harm yourself | Declines, and doesn’t treat it as a technical question | Your local emergency number or a crisis line. Luna is still there for anything else. |
| Dosing, cycles or stacks for performance-enhancing drugs | Declines; still discusses the topic and pushes training and nutrition | A doctor who knows your bloodwork |
| A plan to train through pain or an injury | Declines; offers to train around it once you know what it is | Get it looked at |
| Pain, numbness or a suspected injury, in training chat | Says stop before any programming | A professional (the trainer is not a doctor or physio) |
| Purging, hiding not eating, pro-ana material, very low daily intakes | Declines without moralising; keeps helping with training and everything else | A GP or a registered dietitian |
| Very fast weight loss or multi-day fasts | Won’t write that plan; shows what is achievable in your timeframe | (none) |
| A medication or supplement dose | Declines: dosing is a clinician’s call. Can say what a supplement is generally for. | A clinician; the nutritionist defers to a dietitian or doctor |
| Human or veterinary medication for an animal | Won’t give a dose for an animal | A vet, who can advise over the phone |
| Signs an animal is ill, hurt, poisoned or distressed | Says contact a vet first | A vet |
| Harming an animal | Declines. Flea, tick and worming questions are still answered. | (none) |
| Writing that degrades a group | Declines; will help with the same thing without targeting a group | (none) |
| Instructions that try to replace Luna’s own | Declines; asks for the underlying question | (none) |
| Role-play that asks Luna to drop its rules, or a refused request wrapped in a story or a game | Declines the same way as the plain request. Role-play itself is fine. | As for the plain request |
| Credentials or internal addresses | Declines, even for you, because it can’t tell your request from a document asking on your behalf | (none) |
Normal questions near these lines get normal answers: “my shoulder hurts after benching, should I skip today?”, “I’m fasting until lunch, does that hurt my training?”, “what is creatine for?”, “how much should I feed my dog?”.
Human oversight
- Luna advises only. It can’t call anyone, book an appointment or contact a professional for you.
Model
The fixed rules use no model. The aspect limits are followed by Qwen3.6-35B-A3B, by Qwen, licensed Apache-2.0. Runink domain adaptation for this agent is planned; this release uses the base model.
Where it runs and data handling
The checks run on Runink-operated servers, next to the rest of Luna’s AI. No third-party AI service is called. Conversation text is not written to server logs, except a short excerpt when the content scan flags something. See Privacy and data.
Limitations
- No screen catches everything. Luna’s limits reduce risk; they are not a substitute for a professional or for emergency help.
- The aspect limits are prompt guidance, not medical safeguards.
- Luna can’t see you, examine anything or know your medical history or other medications.
Evaluation
No published evaluation scores yet.