The totals match. Why is the report still on hold?

In a training exercise I run through StarForge Dynamics, we use a fictional source packet with 12 records. Nine records go into the included totals. Three are exceptions. The included budget is $810,000, and included actual spending is $733,550.

The verified workbook has two different answers at the top: reconciliation is PASS, and release is HOLD. That is deliberate.

All numbers and organizations in this exercise are fictional. They are teaching data, not customer finances, savings, or measured business outcomes.

What the arithmetic tells us

Reconciliation asks whether the included records agree with the control totals we expect. It is a useful check. If the included amounts do not add up, we need to find out why.

But a correct sum does not explain every exclusion. It does not tell us whether an unresolved record should have been included, whether the reporting period is right, or whether someone has approved an exception.

Those are separate questions. Putting them in a separate release check makes them harder to overlook.

Keep the missing pieces visible

I would want the report to show which records were left out and the reason for each one. If a required field is missing, preserve that fact. Do not ask the model to make the record look complete.

In this exercise, the three exceptions have identifiable record IDs. A reviewer can go back to the source instead of trying to reverse-engineer a polished summary. The report can say what it knows and leave the unresolved decision visible.

That also makes the next conversation more useful. Instead of asking someone to review an entire workbook with no guidance, we can point to the records that need attention and explain what would resolve them.

Give review a specific job

Before using AI to help with a recurring report, agree on the included population, required fields, and what should happen when a record fails a check. Name the person who can decide whether to correct, exclude, or investigate it.

The release rule depends on the work. Some exceptions can be accepted with an explanation. Others should stop publication. The teaching example uses a HOLD so the class can examine the unresolved records. It is not a rule that every exception in every business report must block release.

For higher-consequence work, the review needs to reflect that consequence. The person approving the report needs the underlying evidence and the authority to make the decision.

Try the small version

The Reconciliation Lab reproduces the teaching distinction with sample values. It shows the included totals and lets you vary an actual-spending total or mark the exceptions reviewed. It is a demonstration, not an audit or a connection to a real workbook.

You can use the same idea without building anything new. Put an exceptions list beside the next report your team prepares. Write down who resolves it and what the report is allowed to claim before that happens.

Basis: a fictional training packet and verified benchmark I built for a StarForge Dynamics power-user course in September 2026. This note explains the exercise. It does not rank models or promise a particular AI result.