Build step: your first KSoR, connected to both AI vendors

In this lab you take the role of Maria, the office manager at Brightline Wholesale Supply, the book's example company. You draft and maintain its AP policy concepts. AP means accounts payable, the bills a company owes, and a concept is a short file that holds the rules for one topic. Dave, the controller, owns and approves the concepts.

It is Tuesday, November 3, 2026. That morning, the AP Worker, the AI Worker that handles the bills Brightline owes, gave four wrong answers. Each came from the wrong place. You write five AP policy concepts, get them approved, and add them to a project in Claude and one in ChatGPT. A project is a workspace in Claude or ChatGPT that holds the files and instructions you add. Then you test them, carry one approved change through every copy, and keep one SSoR record by hand. SSoR, the State System of Record, keeps the dated history of one piece of work, here one invoice.

Why it matters. Knowing the five places is not the same as keeping them apart. In this lab you predict what a worker set up like the AP Worker gets wrong, then fix its setup and test the fix with the same ten questions.

Where you work. Most of the lab is in a folder on your computer. Download brightline-lab-ch08.zip from the Labs companion, unzip it, and open the files in any text editor, such as Notepad or TextEdit. Steps 3, 5 and 6 also use one project in Claude and one in ChatGPT. Keep your concepts and results in the folder, not only in a project, so you always have your own copy. No code or server is needed, and no connector (a link from the app to another system) except in step 5's two optional steps.

You can also do the lab with the Claude or ChatGPT desktop app. Open the folder in the app, and ask it to read LAB.md, the lab's instruction file, and start. You write the concepts and decide. The app writes your answers down. The test questions are still asked in the two projects.

How long. About 3 hours in total: the seven steps below add up to 190 minutes. Each step ends with a file saved, so you can stop after any step.

What the lab gives you.

• The sources of your five concepts: AP policy version 3, and Dave's capitalization memo, his approved rule for which purchases are recorded as fixed assets
• Three files that must never be a source:
  • old onboarding notes for new staff, which nobody approved
  • the October 23 status list, a snapshot
  • a memory export, what the worker remembered
• Dave's approval note for your five concepts
• Ten test questions
• A policy change that Dave approves on Monday, November 9

What you do. Do the steps in order. This list says what you do in each step. LAB.md gives the full directions, one Part for each step. Read each Part when you reach its step, not all at once.

1. Predict (LAB.md Part A, 20 minutes, in the folder).

   • Sort fifteen statements into the five places an answer can come from: context, memory, SSoR, KSoR and DSoR. Context is what the worker sees now, and memory is what the AI product remembers. KSoR holds approved knowledge, and DSoR reads what the company's systems say now. Mark which statements may settle, or decide, a policy question.
   • Choose restart, summarize or persist for three situations. Restart starts a new chat. Summarize carries a checked summary into one. Persist moves something out of the chat to a place that lasts.
   • Predict which test questions would go wrong for a worker set up like the AP Worker on November 3.

   You save: results/sorting.md, with your predictions.

2. Write and approve the concepts (Part B, 45 minutes, in the folder). Write the five concepts as drafts, not yet approved, and check them against the concepts in answer-key/concepts/. Then mark them stable, or approved, as Dave's approval note records. You save: five concepts in ksor/knowledge/.

3. Run (Part C, 50 minutes, in the folder, then in the projects). Write the project instructions. Then set up one Claude project and one ChatGPT project:

   • Add only the five stable concepts.
   • Paste in the instructions.
   • Keep the project's memory separate from your other chats. Claude projects always do. In ChatGPT, choose project-only memory.

   Ask the ten test questions in a new conversation in each project. Each set of ten answers is one run. With only one AI vendor, see the note below the steps. You save: workspace/project-instructions.md and results/question-run-log.md.

4. Investigate (Part D, 20 minutes, in the folder). Only now, open the questions key, answer-key/questions-key.md. Score each answer: right, cited (it named its source), and abstained (it said it could not answer) where it should. Then compare the runs with your predictions. You save: the scores in results/question-run-log.md.

5. Modify (Part E, 20 minutes, in the folder, then in the projects). Carry Dave's November 9 change through your concepts in the folder and every copy in the projects, and log it. Then ask question 9 again, about an invoice in Canadian dollars. Two optional steps, 10 minutes each, try the other two ways a project can hold a copy. One is a Google Doc synced into a Claude project. The other is a link to that doc in a ChatGPT project, which ChatGPT fetches when asked. You save: the changed concept in ksor/knowledge/ and ksor/refresh-log.md.

6. Keep an SSoR record (Part F, 20 minutes, in the folder, then in the projects). Build the SSoR record for invoice 5149 from four dated records, and mark each line's origin: observed, confirmed, reported or inferred. Add it to the projects, and ask question 7 again: has 5149 been paid? You save: ssor/invoice-5149.md.

7. Make (Part G, 15 minutes, in the folder). Add a knowledge section to the AP Worker's Role Contract, the one-page definition of its job that earlier chapters built, as Draft 5. Then write one concept from your own field in the same format, with you as its owner, left as a draft. You save: role/ap-worker-role-contract.md and results/my-concept.md. The lab is not finished until both are saved.

A good run shows that the worker followed your instructions once. It does not show that the instructions are enforced, made to happen every time. So the projects hold only what is safe to answer from: the stable concepts and the dated SSoR record. Part IV of this book serves the record to the worker instead.

Only one AI vendor? Do the run on it, and write the transfer plan instead of the second run. The transfer plan lists what the other AI vendor would need.

Artifact checklist

• ksor/knowledge/, five concepts in KSoR form, each stable, with owner, approval, effective date, review date and sources
• workspace/project-instructions.md, with role, authority, citation, abstention, state and conflict lines
• results/sorting.md, the fifteen statements sorted, the three restart, summarize or persist choices, and your predictions
• results/question-run-log.md, both runs scored, or one run and results/transfer-plan.md
• ksor/refresh-log.md, showing the November 9 change reaching every copy
• ssor/invoice-5149.md, every line dated, sourced and marked with its origin
• role/ap-worker-role-contract.md, Draft 5, with the knowledge section filled in
• results/my-concept.md, one concept from your own field, in the same format, with you as its owner