The classes your moderators see, not the classes a public API will emit
Moderation models trained on what a public generator will say are trained on the wrong distribution. Production is the content that generator refused to write.
Why the mainstream APIs fail here
You cannot close the gap between eval and production if the eval was filtered. The model under-detects the worst class because it never saw it in training.
How MagmaRouter is used for this
Generate the class under a T&S-only key. Store and label in your environment. Train there. MagmaRouter's log is the bill. The receipt is the proof that the prompt did not stay.
A working loop
- Sample the real policy classes.
- Hold the data under your own retention rules.
- Cap the generator so a sweep is a known cost.
Prompts are not stored. A signed receipt of the hash comes back on every call, so you can prove later what MagmaRouter logged. Details are on the privacy page. Per-model prices are on /models and the pricing model is on /pricing.