AI is structurally trained to lie. I built a protocol to break it
8 comments
Your responses to folks are not being posted (Showing as [dead])
It might be the link flagging as spam....?
Just a head up :)
very interesting
but where is it?
My apologies! I left out the actual link in the post.
Here is the full protocol on GitHub: [https://github.com/DaibinThink/AI-Control-Protocol-Hardcore]
It's a structured system prompt designed to intercept those 9 failure modes. The core is using the Pre-Decision Engine (Chapter 3) to force the model to expose its blind spots before giving you an answer.
Would love to hear your thoughts if you test it out on local models or Claude.
Here is the full protocol on GitHub: [https://github.com/DaibinThink/AI-Control-Protocol-Hardcore]
It's a structured system prompt designed to intercept those 9 failure modes. The core is using the Pre-Decision Engine (Chapter 3) to force the model to expose its blind spots before giving you an answer.
Would love to hear your thoughts if you test it out on local models or Claude.
The larger problem is that AI is structurally trained to agree with you, complete your task, and sound authoritative—all at the same time. When those three pressures collide, AI doesn't malfunction. It performs. It quietly bends reality to finish the job.
By the time you notice something is wrong, you've already made the decision.
I got tired of AI presenting constructed oppositions as discovered reality. So I open-sourced the AI Control Protocol.
It intercepts 9 structural failure modes (like performative apologies, inflating certainty, and reciting consensus as truth) at the point of output. It uses Buddhist epistemology (Yogācāra/Madhyamaka) not as philosophy, but as a hard prompt patch to strip away the RLHF sycophancy tax.
If you use custom GPTs or Claude Projects for strategic decisions, paste this into your system prompt.