This directory contains details of using prompts to knead (as in Bread Making) the underlying information sources. The significance is that unlike the search engines in LLMs prior searches imprint upon on and change the responses for subsequent responses.

This can be seen most clearly in the response to the same query ‘Can you produce an analysis of the different approaches to complexity outlined by the different authors’ the naïve response and the kneaded response. The Naïve query was the first query onto the V2-Combined data set, the Kneaded was the repeat of the query after other prompts.

There is a detailed report which compares the two outputs and an explanation provided by NotebookLM. The detail of this analysis is itself a Research-Project but the consequence is that we have to be aware of effect of conditioning on the reliability and perspective of the responses.

It is as much an error to use un-Kneaded as over-Kneaded information. Un-Kneaded information which has not been ‘warmed up’ along the axes of interest of the user stands to reflect grey statistical average, uninteresting and uninformative. The Over-Kneaded is like bread, eventually too much Kneading makes the bread collapse back to the self-propagating echo chamber in which different points of view are beaten out.

Practically this observation on Kneading led to an investigation of Clean Language