Generate the abuse you have to train on
A T&S model trained only on what a frontier API is willing to emit is trained on the polite subset of abuse. The queue is not polite. The training set cannot be either.
Trust and safety, policy, and ML teams labelled to handle this material.
Why the mainstream APIs fail here
The generator refuses the class you are trying to detect. You undersample the worst cases, the classifier looks good on the eval you were allowed to create, and production is a different distribution.
How Redline is used for this
Generate labelled examples under a key that only the T&S tooling holds. Store them in your own bucket. Redline keeps the bill, not the text. A lifetime cap on that key is the control your counsel can read.
A working loop
- Generate the class you actually moderate, not a euphemism.
- Label in your own store.
- Retrain. The gateway does not keep a copy.
Prompts are not stored. A signed receipt of the hash comes back on every call, so you can prove later what Redline logged. Details are on the privacy page. Per-model prices are on /models and the pricing model is on /pricing.