Guardrails

Guardrails are not a single switch. The first layer is built into the model itself: during fine-tuning, often with human feedback, the maker rewards helpful answers and teaches the model to decline certain requests. A second layer is written in plain language in the system prompt, the hidden instructions that frame every conversation. A third layer is made of separate classifiers that read your request, or the draft reply, and block or rewrite it. Finally, the maker's usage policy sets what customers are allowed to ask for, and it can be enforced by suspending accounts.

Makers add them for reasons that are legal, commercial and sometimes ethical: avoiding harm, avoiding bad press, respecting the laws of many countries at once. Typical triggers include weapons, self-harm, medical dosage, financial advice and copyrighted lyrics. Because the lines are drawn by each company, the same question can be answered by one model and declined by another. The cost is over-refusal: a question about how two medicines interact, a dark scene in a novel or a blunt opinion can be turned away or buried under disclaimers although nothing wrong was asked.

Guardrails also say little about who runs the model. A model with few built-in rules can still be served by a host that adds its own filters, and a very cautious model may be reached through a service that adds nothing. When you compare tools, the useful questions are which rules apply, who applies them, and whether the tool tells you when one of them has been triggered. Researchers also distinguish a refusal, where the model says no, from silent dilution, where it answers but leaves out the useful part, which is harder to notice.

Claude Opus 5.5 on NeverNo is a good example of a maker's guardrails staying in place. It is Anthropic's large reasoning model for careful writing and analysis, and it is anonymized: NeverNo sends the request without your name, account or IP address, but Anthropic still reads the text, applies its own usage rules and may keep the request for a while under its own policy. Opus may therefore decline some requests. When a request its maker does not allow comes up, one of NeverNo's private models answers it instead, and the app tells you so.

NeverNo's own models and the open models in the catalog are labeled private, and they are chosen and configured to answer frankly on any lawful topic, which is what the uncensored chat page describes. Three hard limits still apply to every account, model and language, and they are checked on our servers: nothing involving minors, nothing about real people without their consent, and no real help to seriously harm others.