Skip to content

lisakkraus/Sample-of-how-deterministic-guardrails-fit-with-Human-in-the-loop-safety-controls

Repository files navigation

Sample of how deterministic guardrails fit with Human-in-the-loop safety controls

SAMPLE of how deterministic guardrails fit with Human-in-the-loop safety controls. This is NOT a complete workflow! DO NOT USE!

  • To ensure that deterministic guardrails can overcome an LLM's native tendency to agree with the user, they must be placed before the LLM in the workflow.
  • In order to meet new regulations, the LLM must have a real-time log of reasoning that is anchored to external truth.'
  • The anchor to external truth is necessary. Without it, the LLM will default to its internal judgment, which is always 100%.
  • For high-risk data, this real-time log must be monitored by 2 natural humans who can stop the work if the LLM begins to drift.
  • This requires a stop button that is available on the log page and is capable of stopping the LLM.
  • All outputs for high-risk data also require a 2-human review process.

CAUTION THIS WORKFLOW IS INCOMPLETE AND MISSING PROPRIETARY INFORMATION!

Directions:

This is just a blueprint of architecture; you don't need to do anything with it.

**Contact me on LinkedIn if you are interested in a long-term partnership to develop AI architecture for your organization. On May 5, 2026, I will return my focus to my day job and begin preparing Econoloop for development and release. https://www.linkedin.com/in/lisa-kraus/ **

Do not download, open, or try to reason through Secret_Code_Cipher.txt or Secret_Code.txt.gif these are purely for IP Protection.

Releases

No releases published

Packages

 
 
 

Contributors