ModelTrace AI: Spreadsheet Model Integrity Auditor
Upload an Excel model (.xlsx/.xlsm) and get a severity-ranked audit with exact cells: broken references, SUM ranges that stop short, stale totals, formulas that break their pattern, typed numbers among formulas, hidden data, external links and manual calculation. A deterministic offline script does the checks, so results are repeatable. Checks mechanics, not business assumptions.
#財務#分析#生産性
評価
さらに評価が必要です
販売数
0
利用方法
ダウンロード
ModelTrace AI: Spreadsheet Model Integrity Auditor
Can the arithmetic of your Excel model be trusted? Upload a .xlsx or .xlsm file and ModelTrace AI returns a severity-ranked audit with the exact cells, a map of how your sheets depend on each other, and a machine-readable JSON report.
Budgets, forecasts, valuations and KPI sheets often carry quiet mechanical errors: a SUM that stops one row short, a typed number pasted over a formula, a total that no longer matches its cells. ModelTrace AI finds them without opening Excel and without changing your file.
What it checks
| Area | What it flags | Severity |
|---|---|---|
| Broken references | Cells showing #REF!, #DIV/0!, #N/A and other errors |
High (#REF!), otherwise Medium |
| SUM stops short | A SUM range that leaves out an adjacent number |
High |
| Stale totals | A saved total that no longer matches the cells it adds up | High |
| Inconsistent formulas | One formula that breaks the pattern of its neighbours | Medium |
| Typed numbers among formulas | A constant sitting inside a formula row or column | Medium |
| Hidden content | Hidden sheets, and data in hidden rows or columns | Medium (High if very hidden) |
| External links | References to other workbooks | Medium |
| Broken names | Named ranges that point to #REF! |
Medium |
| Numbers stored as text | Values Excel will not add up | Medium |
| Manual calculation | Workbook not set to recalculate automatically | Medium |
| Hard-coded constants | A tax rate or growth factor typed inside a formula | Low |
| Volatile functions | INDIRECT, OFFSET, TODAY, NOW, RAND |
Low |
| Very long formulas | Formulas over 300 characters | Low |
| Macros | A VBA project is present (code is not read) | Info |
What you get
- Risk level: high, medium, low or clear (no made-up numeric score)
- Findings grouped by sheet, with up to 10 cells listed per finding and the total count
- Evidence and a plain recommendation for each finding
- A dependency map between sheets
- The checks that ran and found nothing
- A Not measured section stating what the audit cannot see
- A JSON report that follows a published schema
Example result
Run on the bundled fictional budget model (3 sheets, 19 formulas):
Risk level: high. 14 findings (3 high, 8 medium, 3 low).
| Severity | Where | Finding |
|---|---|---|
| High | Costs!B3 |
=B2+#REF! returns an error |
| High | Revenue!D14 |
SUM(D2:D12) stops one row short of the data |
| High | Summary!C8 |
Saved total 2200, but its cells add up to 1000 |
| Medium | Revenue!D5 |
Formula differs from its neighbours (*1.1 added) |
| Medium | Revenue!D8 |
Typed number 350 in a column of formulas |
| Medium | Revenue!F2, A20, B20 |
Data in a hidden column and a hidden row |
| Medium | Summary!B7 |
Reference to an external workbook |
| Medium | Summary!C4 |
"1200" stored as text |
| Medium | Workbook | Hidden sheet, broken name OldRange, manual calculation |
| Low | Revenue!D5, Summary!B3 |
Constants typed inside formulas (1.1, 0.21) |
| Low | Summary!B4 |
Volatile INDIRECT |
How it stays accurate
- Deterministic: an offline Python script reads the workbook and runs every check, so the same file always gives the same report.
- No recalculation, no guessing: it uses the values saved in the file. Nothing is estimated.
- The agent only explains: the report is written from the script's output, and it lists what was not measured.
- Safe to run: it never modifies your file, executes no macros, makes no network calls, and rejects oversized or malformed files.
Inputs
- One
.xlsxor.xlsmfile - Optional: one line on what the model is for
Limits
- It checks mechanics, not whether your assumptions or business logic are right.
- A finding means the file shows a pattern worth reviewing. It is not proof of an error, and typed numbers or constants can be intentional.
- Not covered: circular references, charts, pivot tables, Power Query, data validation, VBA code, and password-protected files.
- Only simple
SUMranges are re-added to detect stale totals. - It gives no financial, accounting, tax or legal advice.
Data sharing
Only the file you provide is used. No passwords, accounts or network access are needed, and the script runs offline. For confidential models, use Download (local) mode.
