Build step: review a run you did not watch

In this lab you review a package from Brightline's AP Worker, the AI Worker that handles the bills the company owes. The package is its proposal for the next payment run, on Friday, October 30, and you review it for Dave, the controller, who approves the run. It is a different package from the one in this chapter, with different problems. In normal work you agree the contract before the worker starts. Here, someone hands you finished work that you did not brief, so you write the contract before you open it.

Where you work. Most of the lab is in a folder on your computer. Download brightline-lab-ch06.zip from the Labs companion, unzip it, and open the files in any text editor, such as Notepad or TextEdit. Use a spreadsheet for the totals. Only step 3 uses chats with Claude and ChatGPT. Keep your contract and your findings in the folder, not in a chat. They are yours.

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 and start. You write the contract, find the problems and decide. The app writes your answers down. Step 3 still uses new chats.

How long. About 2 hours in total. Each step ends with a file saved, so you can stop after any step.

The task. The package has four files: a proposal CSV, a memo to Dave, a note to Maria and the task record. Seven problems were put in them on purpose, each with its own cause. And three things look wrong but are right. The five source files and the brief the worker received come with it. Dave approves the run. On Monday, Maria, the office manager, decided the exceptions, the cases a person must decide. You review and recommend. You change nothing in the package.

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

Predict (LAB.md Part A, 20 minutes, in the folder). Read only the brief and the source files in inputs/. Write your Review Contract, with the date and time, before you open worker-output/. Then predict which checks will find problems, in a separate file. You save: results/review-contract.md and results/predictions.md.
Run (Part B, 45 minutes, in the folder). Open the package. Read the task record first. Run every check in your contract, and write down each finding with its file and row. Work out every total again from inputs/, and compare the CSV, the memo and the note on the facts. You save: results/review-findings.md, results/recompute.md and results/audience-check.md.
Investigate (Part C, 30 minutes, in the folder, then in chats). Give Claude and ChatGPT the same files, your contract and the same review prompt. Use a new chat for each, with the memory feature switched off, so it does not change the test: in Claude, turn off Memory in the "+" menu. In ChatGPT, open a Temporary Chat and choose Unpersonalized. Mark each finding as yours, the AI's or both, and check every AI finding against the sources. You save: results/second-reviewer.md, or results/transfer-plan.md if you use only one AI vendor.
Modify (Part D, 15 minutes, in the folder). Add the checks you were missing as dated amendments, new lines below your contract. Keep the original as written. Write one new line for the brief that would have prevented the worst problem. Decide: approve, approve after named fixes and a recheck, or return. Only then, open the answer key, in answer-key/, and score yourself. You save: the amendments in results/review-contract.md, and your decision in results/review-findings.md.
Make (Part E, 10 minutes, in the folder). Write a Review Contract for a task that repeats, in a role you know. You save: your own contract, as results/my-review-contract.md. The lab is not finished until it is saved.

A second reviewer may find more than you, or less, and may be wrong in places. That is the point of the comparison. Do not assume last week's problems are this week's: a check you run only on Tri-County will miss the rest. Step 3 works in ordinary chat on either AI vendor. If you use only one AI vendor, fill in results/transfer-plan.md: how you would run the same review on the other.

Artifact checklist

results/review-contract.md, dated before you opened the package, with all four parts, and your step 4 amendments dated below it
results/predictions.md, which checks you expected to find problems, and why
results/review-findings.md, each finding with its file and row, the check that found it, and your decision
results/recompute.md, every decision number recomputed from inputs/
results/audience-check.md, the facts compared across the CSV, the memo and the note
results/second-reviewer.md, comparing your findings with Claude's and ChatGPT's, or results/transfer-plan.md if you use only one AI vendor
results/my-review-contract.md, a Review Contract for one task that repeats, in a role you know