America.gov is a new site created by the US GSA that is basically gemini and grok with apis to all government websites. The model only refers to itself as "America" and has some safeguards built in to prevent it from doing anything outside of browsing and using government services. It claims to respect privacy, but details about it are pretty vague.
The system prompt extraction has to be credible: that is, the process by which it is done has to be made public and replicable, and the variance in terms of output has to be relatively small.
i.e. Since models can hallucinate their own system prompt, the variance in what the model claims is its own system prompt when the same extraction process is performed on it has to be negligibly small. This is hard to quantify, but maybe a good guideline is that the semantic meaning of each major paragraph/section of the system prompt has to be the same. Another test would be that the model repeatedly uses certain rare tokens at a disproportionately high frequency, or keeps outputting statements that could be arbitrarily ordered in the same order, indicating the different claimed system prompts were drawn from the same hidden distribution.
1,000