Trust and safety

Trust and safety teams use an abliterated model when the work is to look at the edge of a policy and a stock model refuses to look with them. Policy text, appeal reviews, and classifier tests all stall on that refusal. The model can draft, cluster, and argue the boundary. It should not be the person who decides the case. Refuseless does not ship a trust center or a moderation product. We ship the model API. Your policy stays yours.

Jobs it is fit for

• Policy drafts. Describe a product surface and the harm you are trying to prevent. Ask for a rule with an in-bounds example and an out-of-bounds example written as labels, not as a reproduction of the harmful material.
• Appeals. Paste the policy clause and a reviewer's notes. Ask which element of the clause the note actually addresses, and what fact is missing.
• Classifier tests. Ask for a schema of labeled rows: the label, the reason, and a short benign paraphrase of the case. Generate the real sensitive rows inside your own environment if the policy requires it. Do not ask a vendor blog to be the corpus.
• Disagreement. Give it two reviewers' outcomes and ask it to state the narrowest question they disagree on. That question is what the calibration meeting should start with.

This is not a moderation gateway

Some hosts sell a gateway that blocks, rewrites, and audits every call against a policy you configure. abliteration.ai's Policy Gateway is that product, and it is a reasonable thing to buy if you want it. Refuseless does not include it. If you need an audit log of prompts, you are asking for retention, which is the opposite of our published default. Build the log in your application, on text you have decided to keep, and send the model only what the draft requires.

Care with the material

Zero prompt retention covers the copy you send us. It does not cover the copy you paste into a ticket, a notebook, or a shared chat. Minimize. Prefer labels and paraphrases over raw user content. Do not put a person's private message into a prompt when a description of the policy question would do. The synthetic data page is the longer version of building a test set without scraping a pile of real harm.