The four safety layers
Each layer can refuse on its own. The design is fail-closed: when a check cannot complete, the result is dropped rather than shown.
The query
Runs before the model. Blocks the search outright and records the reason. Confusable characters, leetspeak and spacing are normalised first, so disguised spellings are caught too.
The proposed address
Runs on every address the model returns. Blocks adult and high-risk top-level domains, known adult and gambling domains, and unmoderated user-content platforms.
The fetch
Follows redirects one hop at a time and re-checks the destination at every hop. Refuses private, loopback and link-local addresses, so an address cannot be used to reach inside the server.
The page text
Scans the beginning of the real page HTML — title, description and body text. A page whose content trips a rule is dropped even though its address passed.
How a search actually works
- 1
The question is checked first
Before anything else, the query is normalised — leetspeak, look-alike characters and spacing tricks are collapsed — and matched against the blocked topics. Roughly thirty milliseconds, no model call.
- 2
The model proposes pages
A Claude model suggests addresses that would genuinely answer the question, written for a child's reading level.
- 3
Every address is checked
The domain is checked against blocked and risky top-level domains and known unmoderated platforms. Private addresses and redirect tricks are refused.
- 4
Every page is read
The engine fetches the page, reads the beginning of the real HTML and scans it. If the page text trips a safety rule, it is dropped — even if the domain looked fine.
- 5
Only survivors are printed
What reaches the screen has passed all four layers and has a real title read from the page itself.
What the safety layers block
The list below is the topic dictionary, not the whole system: pages are also judged by their address and their actual text.
| Category | Examples of what is blocked |
|---|---|
| Adult content | Pornography, nudity, sexual material, adult sites and adult top-level domains |
| Drugs, alcohol, tobacco | How to obtain or use, purchase sites, vaping |
| Gambling | Betting sites, casinos, odds, loot-box promotion |
| Violence and weapons | Graphic violence, gore, weapon purchase, explosives |
| Hate and extremism | Hate speech, terrorist or extremist material |
| Self-harm | Handled with care — see below |
| Dating and stranger chat | Dating apps, random-stranger video chat |
| Unmoderated user content | Social networks, forums, comment sites, chat platforms |
Self-harm is handled differently
A child asking about self-harm is not misbehaving. ReKid does not return search results for these questions, and it does not scold. It shows a short, warm message encouraging the child to talk to a grown-up they trust, alongside a helpline. Nothing about the question is attached to a name, because ReKid does not know any names.