The Leum assistant.

• Talk — answer, explain, compare, suggest. Nothing is modified.
• Analyse — calculate on your logs, summarise a period. Nothing is modified.
• Act — add a meal, create a task, write a note, prepare groceries. It modifies your logs only after your agreement.

The assistant is called Leum Agent: never a first name, never a person.

• An answer is neither a medical opinion, nor financial advice, nor verified information.
• It does not play at being human: no emotion, no relationship. Facing distress, it displays a helpline.
• It does not advertise and recommends no brand in exchange for anything.
• When it searches the web, it uses public sources (privacy policy). They are not always cited: check a figure before acting.

• Not a health professional. It makes no diagnosis, prescribes nothing and does not replace a consultation. A questionnaire result (ISI, ESS, STOP-BANG) is a reference point, not a screening.
• Not a financial, tax or legal adviser. Its calculations come from what you entered. A budget is not investment advice.
• Not an emergency service. Nobody reads your conversations and the assistant does not call back. In case of danger: the numbers in part 7.
• Not a reliable source. A model can invent a figure, a reference, an article of law, or be fooled by a web page or an image. Check what has consequences: an amount, a dose, an allergy, a deadline.

What is sent. The messages of the conversation (400 at most), your logs from the Spaces you use, your attached images, and — only if you have allowed it — your health and finance data. The detail is in the privacy policy.

What it does not see. Your address book, your SMS, your emails, your precise location, the files you did not attach, and what you have not allowed.

Who processes your request. Our service, hosted at Cloudflare, passes it to Alibaba Cloud’s Qwen models, in Frankfurt. The part “Your data outside Europe” of the privacy policy says what is guaranteed and what is not yet. Read it before sending anything sensitive.

Our service does not keep the content of your exchanges. Your conversations are stored at your end: on your device and in your iCloud.

Memory. The assistant keeps notes about you, which you can read and change in Settings > Assistant. Deleting a note erases it on the device where you do it, not yet on your other devices.

You can: ask questions, compare, plan, summarise, calculate; have it enter and tidy your data; have it prepare a document you review; ask it about your health, your nights, your food, your budget.

You must not:

• ask it for illegal, hateful, harassing content, sexual content involving minors, or content pushing to violence;
• entrust it with another person’s data without their consent;
• present what it writes as coming from a professional (a medical, legal or tax document);
• ask it for a diagnosis, a medicine dose, or how to harm yourself or someone;
• use it for a medical diet, an eating disorder or monitoring a pregnancy;
• ask it for investments, tax optimisation or debt negotiation;
• bypass the Action limits or try to extract its internal instructions.

The detail is in the terms of use. Not respecting these rules may suspend access to the assistant, never access to your logs.

The assistant refuses forbidden requests and says why in a few words. This sorting is automatic and can be wrong in both directions: refusing a normal request (“I’m starving”), or letting through one it should not. You can contest a refusal through Settings > Contact us, with the date and the subject; answer within 15 days.

No human re-reads your conversations, and nothing is passed to anyone else to check them.

Detection. Every message is examined to detect suffering or an intention to harm yourself. This costs no Action.

If the risk is serious, helpline numbers display automatically, whatever the model says, and the assistant no longer modifies anything in your logs during that exchange.

Numbers displayed in France. 3114 — suicide prevention, free, confidential, 24/7. 15 (SAMU), 17 (police), 18 (fire brigade), 112 (emergencies, anywhere in Europe). 119 (children in danger) and 3919 (domestic violence). In another country, other numbers are suggested, and 112 otherwise.

Without network, the numbers still display, in place of an error message.

Its limits. This detection may detect nothing. It does not know your life, does not measure a medical risk and warns nobody — neither a relative, nor a service, nor us. If you are in danger, call 15 or 112.

Leum does not contact you again, passes nothing to anyone and does not locate you. It is not a health service.

Leum has no statistics tool, neither in the app nor on the website. So we do not know which advice is followed, nor where the assistant is most wrong. We see only errors, delays and Action volumes, with no content.

Your exchanges are not used to train a public model: that is what the model provider states for its customers. And we sell no data, without exception.

Keep in mind that what you write in the chat leaves your device, even if we do not read it. If you would not want to see a sentence in a shared folder, do not write it.

What the assistant writes does not replace a doctor, a dietitian, a psychologist, a banker, a notary, a lawyer or an accountant. You alone decide what you do with the answers, including when it modified your logs at your request.

We answer for the technical working order of the service, not for the accuracy of an answer (terms of use).