A clean total is nota sound model.

No — and that's the gap it fills. It audits the formula layer, not the values. A totals check confirms the columns add up; it will happily certify a grand total that's built on a #REF! three columns over, or on a tax rate hardcoded inside a formula. This reads the actual formulas and grades them for the faults that silently corrupt a model: error cells and broken external links (structural, dispositive), plus hardcoded magic numbers, copy-paste inconsistencies, and volatile/external dependencies (graduated, flag for review). If you want a self-checking close that grades tie-out math, that's the Finance & Reporting Automation Kit — this checks whether the formulas themselves are trustworthy.

Structural faults are dispositive; graduated signals only flag. A structural fault — an error value (#REF!, #DIV/0!, #VALUE!, #NAME?, #N/A, #NUM!, #NULL!) in a formula or its cached result, or a broken external link — makes the cell BROKEN and forces the whole model BROKEN, because a number built on it is already wrong. A graduated signal — a hardcoded constant that should be an input, a formula whose shape differs from the uniform block it sits in (an overtype), or a volatile/external dependency (NOW, TODAY, RAND, OFFSET, INDIRECT, or an external workbook link) — marks the cell REVIEW and lowers its score, telling you what to clean up without breaking the model.

Each formula cell scores 0–100 on three weighted levers: no hardcoded constant (40), consistent with its row/column block (35), and no volatile or external dependency (25). A cell is SOUND at 80 or above with no fault, REVIEW on any graduated signal (or a score below 80), and BROKEN on any structural fault. The model's integrity score is the mean of the cell scores — context only. Fix-first points you at the single thing to address: the first structural cell in scan order if there is one, otherwise the lowest-scoring non-SOUND cell.

It reads your real .xlsx — it parses the actual formulas, not just the cached values — but it never edits, repairs, or writes back to your file. It flags the faults and points you to the exact cell; the included Fix-It Runbook tells you how to resolve each one. It's deterministic and offline: no AI, no network, nothing uploaded, so the same file always produces the same verdict. The companion workbook reproduces the same per-cell and model logic on a pasted sample so you can see the scoring without running Python.

No. It's a structural integrity check on a file's formulas, not an accounting or audit opinion — it asserts nothing about whether your numbers are correct, only whether the formulas producing them are trustworthy. It never scores or ranks people. Confirm any figure that actually matters with the person who owns the model. It pairs with the CSV / Data-Export Schema & Quality Gate (is the data going in structurally sound?), the Document Field Validator (are the records valid?), and the Finance & Reporting Automation Kit (does the close tie out?).