Use this independent review protocol before applying an AI repair. The linked research transcript contains an unsafe formula suggestion; its historical score is not a validation of a working spreadsheet.
Work on a copy. Preserve the original file, calculation mode, formulas, named ranges, external links, and visible values before accepting any repair. An AI report is a list of hypotheses; the spreadsheet application and the declared business rules are the test environment.
The plain-text Markdown checklist includes a blank decision log, a worked example, and a run-record template. Open it in any text editor; it does not execute formulas.
A valid-looking formula can still be wrong
This separate teaching example uses B2:F2 values of 100, 200, 300, 400, and 500. The declared rule is to total all five values in G2. Adding the source values independently gives 1,500. These are illustrative inputs, not a new AI benchmark or an executed workbook.
| Proposed G2 formula | Check | Decision |
|---|---|---|
=SUM(B2:E2) | Totals 1,000 and omits F2 | Reject: incomplete range |
=SUM(C2:G2) | Includes G2 itself and omits B2 | Reject: circular reference and incomplete range |
=SUM(B2:F2) | Matches the declared five inputs and the 1,500 control total | Candidate: verify in the workbook before accepting |
Inspect the cell address as well as the formula text. See Microsoft's circular-reference guide. Keep iterative calculation off for this example; enabling it does not repair an unintended loop.
The five-check formula-audit protocol
| Check | Pass condition | Failure it catches |
|---|---|---|
| Error inventory | Every error value, warning, and unresolved external reference has a cell address | An obvious #REF! is mentioned without a complete workbook inventory |
| Formula pattern | Each copied formula is compared with the rows and columns around it | A syntactically valid formula skips or repeats a row |
| Lineage and rules | Inputs, ranges, signs, units, and business rules are traced to their source | A plausible repair implements the wrong calculation |
| Independent total | Critical outputs are recalculated from source values outside the repaired formula | The workbook and the AI repeat the same hidden error |
| Acceptance log | Every proposed edit is accepted, rejected, or held with evidence and an owner | Suggestions are applied without a reviewer-ready decision trail |
1. Build a cell-addressed error inventory
Record the workbook, worksheet, cell, current formula, displayed value, error type, and the evidence that made the cell suspicious. Include errors that Excel can detect as well as formulas that calculate without an error code. Microsoft's Formula Auditing tools can expose error values, watched cells, and step-by-step evaluation, but they cannot decide whether a valid formula matches the organization's intended rule.
See Microsoft's current formula-error checking guide for the application controls. Keep the resulting inventory beside the AI report rather than replacing it.
2. Compare formulas as patterns, not isolated strings
Show formulas and compare each flagged cell with the corresponding formulas above, below, and on adjacent sheets. Normalize relative references before comparison so that an expected row shift is not mistaken for a defect. Flag unexpected constants, broken references, omitted periods, and range boundaries that do not move with the surrounding pattern.
3. Trace every proposed repair to its inputs and rule
For each repair, record the precedent cells, expected unit, sign convention, time period, and the written rule the formula is meant to implement. A formula can be internally consistent and still be wrong for the report. Hold the edit when the business rule or source field is missing instead of allowing the AI to infer one.
4. Recalculate critical outputs independently
Rebuild control totals from the smallest trusted source range, or use a second calculation method that does not depend on the proposed formula. Recalculate after each accepted change, then compare the result with the original baseline and the expected accounting or reporting identity. A changed number is not proof of a corrected number.
5. Require an acceptance marker for every edit
A reviewer-ready log needs the cell address, original formula, proposed formula, reason, supporting source, independent check, decision, reviewer, and timestamp. Unresolved external links, volatile functions, circular calculations, and undocumented business logic remain exceptions. Never overwrite the only copy of a workbook or silently accept a batch repair.
Why the archived AI repair should be rejected
WFT-009 proposes SUM(C10:G10) in G10. That range includes G10 itself, creating a circular reference. The source does not contain the full workbook needed to establish the correct replacement range. Reject that proposal and hold the repair until the actual inputs and intended rule are available.
The archived 8/10 score describes the original static review, which did not establish formula correctness. This September 13, 2026 editorial check identifies the self-reference without rewriting the immutable research transcript or claiming to have recalculated its workbook.
What the paired test did—and did not—verify
The fictional WFT-009 test supplied a monthly-reporting scenario with a #REF!, a copied range error, a reversed sign, an omitted month, and six control values. The first text response passed three of five static checks. One failure-only correction reached 8/10, but task fidelity still failed. No workbook was generated or recalculated, so the proposed repaired total was not independently validated as a live spreadsheet outcome.
Boundary: an audit is not approval to use the workbook
This protocol validates a bounded review trail. It does not certify financial statements, tax work, regulatory reports, or operational decisions. A qualified workbook owner must define the business rules, review material exceptions, run the file in its intended spreadsheet version, and approve any production change.