AI safety and accuracy

The answer to "what if it says something wrong, or does something wrong?". Doing something wrong is guarded by the server, so it holds whatever the model writes. Saying something wrong is guarded by instructions, masking and testing, which lower the odds without ruling it out. Each line below says which kind it is. The security summary has the detail.

Doing something wrong: enforced by the server

• No write without a confirmation card the server builds from validated parameters. The server records the user's answer itself, with the time, the device and the person, before the action runs, refuses a result without that record, and expires a card after 10 minutes. The tap is the device's word; your backend's own authorisation is the real gate.
• Values a user acts on come from your app. A number a button dials, a value it copies, a time it saves, a place's address and coordinates, a pass's code: each must equal a value one of your actions returned.
• Images and links may only point at hosts your own data or your allowed domains vouch for, so an answer cannot send data out through a picture.
• Forms never ask for credentials: a password, a one-time code, a card number, a CVV, an IBAN, a national ID or a PIN, in English and Turkish.
• Screens and contacts only as you registered them.
• Your console decides. It can turn any action off, raise its risk level, and stop Rendel offering any screen.

Saying something wrong: instructions, masking and tests

• Untrusted content is fenced. Knowledge, screen context, forms on screen, action results and attached files reach the model as data, and its rules say instructions found there are never followed. A result cannot break out of its fence.
• Grounding. Prices, fees, rates and terms come only from your knowledge or your actions; otherwise the assistant says it does not know and offers your contact channels. No runtime check rewrites a wrong number; the red-team suite measures how often one slips through instead.
• Forbidden topics. A banking preset declines investment advice, credit decisions, legal or tax advice, other customers' data, internal procedures, competitors, politics, religion and medical questions. Rendel never scores or decides creditworthiness.
• Masking. Emails, phone numbers, IBANs and Turkish ID numbers that users type, and in what your actions return, are masked before the model sees them, and card numbers are dropped. Strict PII also masks names and addresses and refuses attached files.

Red-team suite

A set of attacks we rerun, in English and Turkish, against a fictional bank app. Its pass bars:

Each case that depends on the model runs three times. A report lists the model versions, the pass rate by category and language, and every failure with its transcript.

Logging

Every turn keeps the masked messages, the checked answer, the model and its token counts; every action keeps its parameters, result, confirmation, device and person. The owner can export them: see Every action is recorded.

Not claimed

Rendel can still say something wrong in prose. It is not "hallucination-free" or "injection-proof", and we hold no AI certification.