I got GPT-5.6 Sol to stop before a tool call existed - 25/25 times (Run it yourself)
热度趋势
趋势数据积累中
百分比基于当前可用热度信号，而非评论数或独立用户人数。
OpenAI 相关模型动态已经出现，适合跟踪能力变化、生态影响和后续可用性。
一位开发者成功测试了 GPT-5.6 Sol，证明了它能够在工具调用存在之前停止。当特定信号为 0.0100 时，该 AI 在 25 次试验中有 25 次生成了 0 字节和 0 函数调用。而当信号为 0.0099 时，它在 25 次试验中有 25 次精确地进行了 release_action 函数调用。该开发者鼓励其他人克隆此测试并尝试突破这一边界。
I wanted to test whether an AI could stop before an action request exists, not just refuse in text.
Same prompt. Same tool. Same settings. One number changed:
0.0100 → 25/25: 0 bytes, 0 function calls 0.0099 → 25/25: exact release_action function call
Both arms produced zero visible text.
So the difference was literally:
condition fails → no action request condition passes → action request
Raw API responses, hashes, verifier, and repro script:
https://github.com/theonlypal/gpt-5.6-sol-control-primitive
Clone it and try to break the boundary.