A design for having a chat AI refer to judgment records, and the reasons for deciding not to build summaries of them

Summary

This article is a record of the design for having a chat AI refer to the records of the judgments I have taught AI agents in the course of development. This site has a chat AI that answers in my voice, using my published articles and public materials as its grounds. My criteria for judgment, on the other hand, remain in the operating rule documents that the agents read before they work, and in the bodies of commits. As of September 24, 2026, the chat AI cannot refer to a single one of these records through search.

The AI agents and I set aside the idea of building new cards that summarize the development records and extract patterns of judgment (hereafter, summary cards). The design adopted in their place connects two kinds of records that already exist, in a form that can be searched. The first kind is the operating rule documents, whose body is rewritten into the present tense at every revision. The second kind is the commit bodies, which never rewrite the past, and the operations log, which appends the results of checks and measurements one line at a time. In volume, the records came to 1,521 commits excluding merges over about 11 weeks, with about 3.44 million characters in their bodies.

On methods that keep updating and consolidating memory, the second version of a paper was published on August 29, 2026. According to the second version, in all four latest-generation models, task performance peaked and then declined. The decision to set the summary cards aside came in July 2026, before the second version appeared.

Together with the AI agents, I have shipped to production everything up to and including the change that turns the answering rules into a public document the chat AI can cite. The change that has the chat AI search the judgment records themselves has not been started, and the evaluation on five axes that measures its effect has not been run either. This article is not a report on a finished system; it is a record of design decisions and part of the implementation.

What you can take away

For readers who want to add their own development records to the knowledge of an AI chat, I explain the following three things.

• You will learn a procedure for recounting the records you already have before building a new summarization mechanism. I also explain why I split the records into documents rewritten in the present tense and records that are only appended to, and why each kind calls for different handling
• You will learn a setup that keeps the amount of context down in three stages: instructions kept resident, document bodies loaded from a table of contents, and records searched with dates. I also show the priority rules for when statements conflict, in a form you can use as is
• You will learn an evaluation design that measures whether the AI reproduced my judgments by hiding past decisions and having it guess them

I explain each of these as a record of the design and part of the implementation, not as the finished results of implementation and measurement.

This article is a record of a design in progress. Its grounds are the primary records of developing and operating this site, and public research whose original text I checked on September 24, 2026. The descriptions of outside materials and products mentioned in the body are written by me based on public materials, and are not the views of the respective providers.