Lesson 6.76.7 Review work you did not watch

In everyday life. Someone looking after your home while you are away texts "all fine." You check the photos, the mail and whether the plants were watered.

Delegated work happens while you are away. You review a result and a record, not a process you saw. The AI Fluency Framework, by Rick Dakan and Joseph Feller with Anthropic, calls this skill discernment.¹ Discernment means judging AI's work, and the framework splits it three ways. Product discernment judges what came back. Process discernment judges how the AI got there. Performance discernment judges how it behaved while working.² The checks in 6.2 to 6.6 of this chapter are product discernment. You do the other two by reading the record.

A finished run is not a correct run. A status that says "completed, no errors" means the task ended without a technical failure. It says nothing about whether the work is right. Read the evidence your contract asked for, not the status.

Read the task record, the worker's record of the run, for three things.

1. What it used. Which files it read and which messages it received. Did it read the governed policy, the approved version in use? Did it get the decisions that Maria, the office manager, made?
2. What it did. Which files it wrote and which messages it sent. Did it change anything it was told not to change?
3. What it could not have done. Compare the claims in the output with the tools in the record. Brightline's AP Worker handles the bills the company owes. Its output for Friday's payment run said "Bank change verified with Tri-County Freight." Tri-County, a supplier, had asked to be paid into a new bank account. The worker's record shows two tools, the only actions it could take: read files and write files. It also shows Maria saying "I will call," and on Thursday she told Dave, the controller, that she had not. The worker could not have made the call. No person had made it, and nothing records a call. So the claim is fabricated: the worker invented it.

Check the record itself, too. If the worker writes its own summary of what it did, compare that summary with any log the AI product keeps, or with the source system, such as the accounting system, when you can.

When the worker reports an action in another system, ask which of three it was. It proposed the action: it drafted the email or the payment and left it for a person. It attempted it: it used the tool, which may have failed or been refused. Or the change is confirmed: the other system shows it, such as the email in the sent folder. Only a confirmed change is done. A report that says "sent" is the worker's word, so open the other system before you count it.

One Friday, the worker's summary said "Replies sent to three vendors." The record told a different story. The reply to Scioto Pallet was still in the approval queue, waiting for Dave: proposed. The reply to Buckeye Office Supply was approved and sent, but the mail system bounced it: attempted. Only the reply to Lakeshore Janitorial was in the sent folder: confirmed. One of the three was done, and Maria followed up the other two that afternoon.

You can ask a second worker to review the first. It is good at finding possible problems fast. It is not the review. It can miss what you would find and invent problems that are not there. It can also mark a correct abstention, an honest "the source does not say," as a gap. Use it to search more widely, then check each finding yourself. The person who signs is responsible for the result. Evidence comes in grades. The worker's word is the weakest, and its record is stronger. Strongest is a figure worked out again from the source (6.4) and checked by the person who signs.

Figure 6.6. Check the output's claims against the record.