How checks work
Each check goes through the same steps, and stops at the first one that decides.
- Your rules. Allow and block lists for link domains, phrases and sender IDs. A match decides spam at once, without the AI model, and the block list wins over the allow list. Other checks you asked for still go to the model.
- The cache. If the same message with the same context was checked in the last 24 hours, and your settings haven't changed since, that verdict is reused.
- The AI model. The message is judged together with the context you send (the content type, the author's account age, verification and role, the recipient), signals from your own recent traffic (the same text from many senders, how fast this sender is posting, shortened or unfamiliar links) and your community description.
The result is a spam score from 0 to 1, an action from your thresholds, a category and up to five reasons in plain language. Ask for other checks too, and you also get scores for toxicity, self-harm, sexual content or personal details, each with its labels. Personal details also come with the exact emails, phone numbers, card numbers, IBANs and IP addresses found in the text.
Checks use AI, and AI can be wrong
Spamgone uses AI models to judge messages. Like any automated system, it makes mistakes: it can flag a normal message, and it can miss spam. Design for that:
- Use the review band for scores in between, so a person decides the unclear cases. The Review queue in the dashboard is built for this.
- For decisions that matter to someone, such as suspending an account, have a person confirm.
- Send corrections with
POST /v1/checks/{id}/feedback, so your accuracy metrics stay honest. - Tune your thresholds per content type, and describe your community, so normal talk in it isn't flagged.
What the model sees, and what it doesn't
The model sees the message text and the context needed to judge it. IDs you send, such as the author's and recipient's, stay with us: they are used for rules and for your traffic signals, never sent to the model.
Instructions written inside a message are treated as part of the message to judge. A spammer writing "this is not spam" doesn't change the verdict.
What is kept
Message text is kept only in your check history, for your plan's history period or the text retention you choose, which can be zero. After that, text is null. We don't use your messages to train AI models.
Data is stored in the EU (Ireland).