I've been doing a lot of work around this area and the issue tends to pop up around context. There are instances a haiku model will catch something far more accurately than gpt-oss-safeguard. Obviously you pay in costs, realistically, you need a full system here for a proper implementation so some flows to tiny LLM some flows exclusively ML, some flows to a higher intelligence.
People on the fediverse are generally pretty anti ai but it's basically impossible to scale a platform without AI moderation. I would fully welcome any instance trying an implementation of [Osprey](https://discord.com/blog/osprey-open-sourcing-our-rule-engine with LLMs.

Spot on. Most engineers at companies are using LLMs nowadays but when we use it in our daily lives it becomes "vibe coded".
I've started using the term agentically built or agentically driven instead.