What is actually in the reconciliation log?

Each line holds the value in the model, the value in the source document, the name of that document, the page it was read from, the version date, and the result of the comparison. The log for this build runs to 4,375 lines covering 4,350 figures reproduced from IRS Publication 15-T. It is a plain file that opens in the same program as the model.

You used AI. Why should I trust the output?

You should not trust it on the strength of a promise. AI tooling produces arithmetic that looks correct whether or not it is, which is exactly why every figure was compared against the published source and the comparison written down. The reconciliation log is the answer to this question: it names the page each number came from, so any line can be checked without asking the builder anything.

How is this different from a normal spreadsheet audit?

A typical audit is a sample. Someone opens the file, tests a set of cells and reports whether the results look reasonable. This is full coverage instead: every figure that enters the model is compared with the source document and logged individually. The deliverable is not an opinion that the file looks fine, it is a line-by-line record you can read and disagree with.

Can someone else verify the model without me?

Yes. That is the point of shipping the log. A reviewer needs the workbook, the log and the source document. Each line names a page, so checking a figure means opening the source at that page and comparing two values. No knowledge of how the model was built is required.

How do I verify an Excel model I already have?

Start from the source document rather than the spreadsheet. List every constant the model relies on, find each one in the source, and record the model value, the source value, the page and the date side by side. Anything you cannot trace to a page is a figure nobody can defend. This is the same reconciliation pass described above, run against a file that already exists.

Does the model or the log give tax advice?

No. This is an engineering and verification job. The work is reproducing figures from a source document accurately in a spreadsheet, and evidencing that the reproduction is faithful. It offers no guidance on tax, accounting or legal matters, and no view on how anyone should handle their own affairs.

What happens when the source document is revised?

The log records the version date of every figure, so a revised document can be compared against what the model currently holds instead of being rebuilt from memory. The figures that changed are the ones with a new value on the source side; the rest reconcile as before.

Inside a 13,815-formula Excel payroll withholding model

THE BUILD

Nine tabs, 13,815 formulas

The client needed a US federal payroll withholding engine in Excel. Not a summary and not an estimate: a file that takes the inputs and returns the same figures published in IRS Publication 15-T.

Publication 15-T contains percentage-method tables. Those tables are a grid — brackets, thresholds and per-band amounts, repeated across filing statuses and pay periods. Reproducing that grid in a spreadsheet means turning printed rows into lookup tables, then writing formulas that select the right row for the right combination of inputs.

The finished workbook has 9 tabs and 13,815 formulas. It reproduces 4,350 published figures, to the dollar.

The workbook is not the whole deliverable. The evidence that it is right is part of the deliverable.

THE FAILURE MODE

A wrong number does not look wrong

Most spreadsheets fail loudly. A broken reference gives #REF!. Text in a maths operation gives #VALUE!. You see the error and you fix it.

A spreadsheet that decides money fails quietly. Transpose two digits in a threshold and nothing breaks. No cell turns red. The model returns a number, the number is the same shape as every other number in the column, and somebody acts on it.

That is the problem with this class of work. The output of a correct model and the output of a wrong model look identical. You cannot tell them apart by reading the answers.

So the question is never whether the file runs. It is whether every constant inside it matches the document it came from.

THE METHOD

What reconciled line by line actually means

Line by line is a literal description, not a figure of speech. Every figure that enters the model gets its own row in a separate log. The log for this build runs to 4,375 lines.

Each line records six things.

The value as it sits in the model.
The value as printed in the source document.
The name of that source document.
The page the figure was read from.
The date of the version used.
The result of the comparison between the two values.

You do not have to trust the builder. You can check the work.

MISMATCHES

How a wrong figure surfaces

A mismatch is not a judgement call. Two values are compared. They agree or they do not.

When they do not, the line carries the discrepancy and the page reference that settles it. Resolving it means opening the source at the page named on that line. There is no re-derivation, no guessing at what the builder was thinking six hours earlier, no hunting through the workbook for where a constant was typed.

This is why the reference is written down at the moment the figure is entered rather than afterwards. A log assembled at the end records what someone remembers. A log written during the pass records where the number actually came from.

SAMPLING

Why a spot-check is not verification

The usual review is a spot-check. Open the file, test a dozen cells, see whether the answers look sensible. It is quick, and it is better than nothing.

It is the wrong tool here. A spot-check tells you about the cells you checked and nothing else. Errors in a transcribed table are not spread evenly — they cluster where the reading was hardest. The last row before a page break. A band that changes width. A column that repeats four times with one value different. Those are the cells a small sample is least likely to land on.

With 4,350 figures, any sample fast enough to be cheap is too small to count as evidence. Full coverage means checking all of them and leaving a record that shows you did.

The honest answer about AI

AI tooling was used to build this file. That is not a disclaimer at the foot of the page. It is the reason the reconciliation exists.

Language models are very good at producing arithmetic that looks right. They will fill a table with plausible values in the plausible shape, with the right headings and the right formatting. Plausible is a low bar. It is not the standard for a file that decides what someone is paid.

The failure is specific and predictable. A model will supply, with complete confidence, a threshold that was never printed anywhere. It does not know it is guessing. Nothing in the output marks the guess as different from the rest of the column.

Reconciling against the source is what makes AI-assisted work checkable. The tooling speeds up building the structure. The pass against the published document decides what stays. Any figure without a source line does not ship.

That is the argument, and it is deliberately not ‘we use AI carefully’ — a claim nobody outside the room can test. It is: here is the log, here is the page each number came from, check any line you like.

Plausible is not the standard.

WHAT SHIPS

What the client received

Two files, delivered together.

The workbook. 9 tabs, 13,815 formulas, reproducing 4,350 published figures.
The substantiation log. 4,375 lines covering all 4,350 figures, naming source, page and date.
A record of which version of the source the model was built against, so a later revision can be compared with it rather than guessed at.

TRANSFER

The method is not specific to one document

Nothing in the process depends on which document is on the desk. It runs the same way for a rate card, a commission scheme, a freight tariff or a licence-fee schedule — anywhere a printed or PDF source has to become a working model and the figures matter.

The client supplies the source. Every figure in the model is checked against it and recorded. The log ships with the file.

The same method covers an existing spreadsheet. There the model is already built, and the source document decides which of its constants survive.

13,815 formulas, checked one at a time