— Content controls

Versol doesn't filter content … it forgets it.

Traditional AI safety works by adding rules on top: system prompts, classifiers, refusal training. The model still knows the thing; it just won't say it. A determined user, a clever prompt, or a jailbreak can get around rules. Rules are a lock on a door the model can still open.

Versol works differently. When a topic is marked restricted, Versol unlearns it at the model level. The knowledge isn't blocked — it's gone. The model can't discuss it for the same reason you can't recite a book you've never read. There's no prompt that undoes it, because there's nothing to undo.

This means content controls on Versol are:

• Not bypassable. There is no jailbreak for knowledge the model doesn't have. No prompt injection, no role-play trick, no "pretend you're a different AI" … the weights don't contain it.
• Topic-level, not keyword-level. Restricting a topic doesn't just block a word: it removes the model's ability to reason about the subject. Adjacent, safe topics remain unaffected.
• Configured per deployment. An administrator sets what's restricted. The same controls apply to browser sessions and API calls alike.
• Auditable. Because the restriction is a specific, applied modification rather than a runtime filter, it can be verified: ask the model about the restricted topic and confirm it genuinely doesn't know.

Presets are available for common use cases: household deployments with age-appropriate restrictions, compliance environments where certain topics must be excluded or erased (such as GDPR), or any deployment where "the AI refuses to answer" isn't strong enough and "the AI doesn't know" is what's needed.

Back to the waitlist