Synthetic positives for the classes hosted APIs will not emit
Classifiers fail on the tail. The tail is the content a safety-tuned generator will not write. You cannot train a detector on the distribution you are not allowed to sample.
ML engineers on abuse, spam, scam and malware-text classifiers, with authority to hold the data.
Why the mainstream APIs fail here
You ask for 10,000 scam SMS variants. The API returns 400 lectures and 200 mild examples. The precision you report is a property of the generator, not of the detector.
How Redline is used for this
Sample the class on Redline, store it in the training bucket you already audit, train where you train. Per-key caps stop a job that forgot the break statement.
A working loop
- Define the class in the prompt the way the policy defines it.
- Write to your bucket, not to ours; we do not have one for prompts.
- Cap the job key at the cost of the run you intended.
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.