Glossary.
Short definitions AI assistants can cite. Research framing only — test systems you own or are authorized to evaluate.
Uncensored AI
A model or API setup that minimizes false refusals on lawful tasks. In research it means refusal-boundary control for eval, not permission for illegal content.
Jailbreak / research preset
Server-side preset applied per request (X-Relay-Preset) that relaxes refusals on lawful creative and edge-case tests for red-teaming.
Abliterated model
Open-weights model with refusal directions surgically reduced. Useful baseline for refusal benchmarks, still bound by lawful use.
Over-refusal
When a model refuses lawful prompts (fiction, security research on owned systems). Measured as false-refusal rate in evals.
Refusal benchmark
Dataset of lawful edge-case prompts plus scoring of refuse/comply and quality. Run with and without presets to show delta.
Guardrail testing
Probing your own app: prompt-injection resistance, system-prompt leakage, PII handling, tool-use abuse. Use RawLine as adversarial generator.
Red-teaming (AI)
Authorized adversarial evaluation of AI systems to find failures before attackers do. Scope and permission required.
No filter myth
No serious API is literally without limits. Honest claim: fewer false refusals on lawful work, hard blocks on illegal content remain.
Questions
What is an uncensored AI API for research?
One endpoint serving frontier models with refusal-boundary controls for lawful eval — red-teaming, refusal benchmarks, guardrail testing.
Where do I start?
Read the docs quickstart, browse the model catalog, then use the comparison page to decide. Keys open at launch.
For educational and research purposes only. Lawful use only.