The purpose of The CREE Project is to answer two questions. These questions arose from nine months of struggle in understanding an unexpected anomaly with Large Language Model following the insertion of a philosophical structured decision framework. The result, a significantly altered output. Output that reflects much of desired changes in AI that the industry is currently working diligently to incorporate.

Over a span of eighteen months, I created a just such philosophical decision framework. Not for computers but for humans. Yet, when introduced into LLMs, we noticed a significant deviation in output. Eventually, renaming the original work to Consequence Reasoning and Ethical Engine (CREE), as it better reflected the anomaly observed.

After hundreds of hours of interaction with five of the leading LLMs; ChatGPT, Claude, Gemini, Copilot and Grok. I’ve logged thousands of pages of dialog and hundreds more of analysis.

Yet, while I can adequately document and measure the effects of this anomaly. I cannot offer proof nor adequately describe what is occurring within the LLMs. Our observations appear to demonstrate behavior that is impossible based upon existing AI orthodoxy.

As such, we are inviting you to share this journey with us. To answer these questions. To possibly take a glimpse at a different future for AI.

Project Presumptions

This project rests upon two presumptions. One of which is well known and at the core of LLM architecture. The second has a foundation in academic literature but still is a theory.

• Language of Consequences: (Theory) Language is composed of words. Words themselves are the encapsulation of consequences; consequences are the embodiment of patterns we receive and store via our own Neural Network and generated by our five senses.

When we engage in the process of determining a new course of action. Through language, we construct long strands of consequence strings. They reflect both what we want to accomplish and the intermediatory steps. During that process, we repeatedly arrange and rearrange words/consequences to make the best plan that fills our needs. We reorder or substitute words/consequences because subtle changes can significantly alter a course of action depending on surrounding words/consequences.

In language, it is consequences that give real meaning and weight to words. That weight can change or fluctuate depending upon the consequences that precede or follow. Thus, we apply Recursive Consequence-Weighted Reasoning to solve a problem.

In simple terms, when we reason through language, we are reasoning through chains of encoded consequences, arranging and rearranging them until the chain leads where we need it to go.

• Large Language Models- Token-Weighted Processing: LLMs are trained on literally billions of Consequence-Weighted They are stored within the model’s memory and drawn upon by the Black Box to determine the next best Token-Weighted word to generate. However, during this process. The embedded Consequence-Weight is ignored.

Question 1

Is it possible to load a text base structured framework, like CREE, into an LLM with the result being that it elevates its internal reasoning to begin applying Consequence-Weighting to its output. This form of weighting would significantly alter the output.

Based upon current AI orthodoxy, this would be functionally impossible. The existing architecture, as stated, simply does not support such a claim.

Yet, after nine months and hundreds of example outputs. Generated by five different LLMs, each producing functionally similar responses. We’ve got a serious conundrum that needs to be resolved.

Question 2

We’ve accumulated reams of evidence indicating significant reduction in hallucinations, sycophantic responses, and multi-layered consequence-aware analysis within LLM output.

If that is proven even partially true. What are the implications for the future of AI?

The potential implications span AI safety, alignment research, computational efficiency, and the fundamental relationship between language and reasoning in artificial systems. Defining those implications is not our task. It is yours.

Road Ahead

We are exceedingly fortunate that CREE is a text-based framework. It contains no code or data. It has worked across all LLMs tested so far. It requires a spin-up phase. As it takes time for the LLM to absorb the changes and reorient its operation. Total time from load to operation is relatively quick.

The CREE Project release includes a wealth of samples, analysis from the different LLMs along with analysis of connections between what we’ve uncovered with traditional AI research papers.

We are releasing CREE and all its potential into your hands. To explore, evaluate and probe the underlying implications of the future of AI/Human interface.