Radar Intelligence Protocol
Frequently Asked Questions
How EmberGF verifies filter triggers, token boundaries, and model architecture.
Why do companion apps silently tighten filters?
Most silent censorship updates are driven by third-party payment processors (Visa/Mastercard adult merchant tier restrictions) or mobile app store policies (Apple App Store Guideline 1.2 and Google Play UGC safety guidelines). Providers often update model system prompts or safety endpoints without notifying users to avoid app store delisting.
How does token-level guardrail detection work?
Guardrails operate either as a pre-inference regex/classifier or a post-inference streaming filter. When a user input contains blacklisted semantic tokens, platforms like Character.AI intercept the generation before the model outputs text, displaying a canned safety message. Open-weight platforms like Candy AI remove this layer entirely.
What is the difference between open-weights and cloud proxies?
Services using OpenAI, Anthropic, or Google cloud endpoints inherit commercial terms of service that strictly ban NSFW generation. Services running self-hosted open-weights (such as fine-tuned Llama-3 or Mistral instances) maintain complete sovereign control over their prompt filters and cannot be shut down by external cloud APIs.
How are community votes validated?
Readers submit reports directly through the radar cards. Votes are hashed client-side with session-bound persistence to prevent automated spamming. When reports for an uncensored platform exceed a 15% threshold within a 48-hour window, the Ember editorial desk performs an immediate hands-on audit with standardized prompt sets.