A workbook with correct formula text can still show stale cached values if calculation is manual or the file has not been recalculated by a compatible engine. A release gate reopens the edited file, checks calculation settings, forces or confirms recalculation as needed, and independently verifies selected results and totals. It also checks formula errors, unexpected blanks, protected ranges, changed sheet names, and links to external workbooks. The reviewer should compare the final file hash with the file placed in the download or shared location. If the export is CSV, recheck its field types and formula-looking values in the target application. A model explanation is evidence of intent, not of the released workbook's state.
Spreadsheet release checks: recalculate, reconcile, then inspect
Operational case
The depot's new F2 formula is correct, yet the saved workbook still displays 11 from an earlier B2-D2 calculation. The reviewer opens it in the target spreadsheet application, confirms calculation mode, recalculates, and sees 16. A separate row-level calculation verifies RM-58 at 31+0-12=19 and RM-63 at 4+11-2=13. Against thresholds 18 and 15, only RM-47 and RM-63 need reorder review. The team checks F2:F48 for formula errors, confirms E19 remained unchanged, and records the hash of the file users actually receive. If the download hash differs, release stops.
RM-47: 18 + 7 - 9 = 16 < 20: reorder.
RM-58: 31 + 0 - 12 = 19 >= 18: no reorder.
RM-63: 4 + 11 - 2 = 13 < 15: reorder.
Gate: recalculate; inspect F2:F48; compare independent rows.
Gate: approved change list and released file hash match.Performance and operating cost
For N simple formula rows, an independent pass is O(N) time and O(1) working space if rows are streamed. Workbook-engine recalculation cost depends on its dependency graph; it can exceed O(N) when formulas reference large or volatile ranges. Sampling only a few rows is fast but cannot prove all 47 rows are sound, so scan the full edited range for error states and independently recompute business-critical aggregates or a chosen set of cases. A full application round trip is slower than checking prompt text, but it catches cached-value and import behavior that the model cannot observe.
Common Mistakes
- Do not release a workbook based only on formula strings or old displayed values.
- Do not assume every spreadsheet library evaluates formulas when saving.
- Do not review one file and distribute another.
Connected lessons
- Production prompt engineering
- Prompt Engineering
- Workbook edits: review a cell diff before applying changes
- Aggregation prompts: pin the denominator and recompute the rate
- Prompt release artifacts: version the whole decision path
- Workbook prompts: name sheets, ranges, types, and provenance
- Formula prompts: test references, types, and fill behavior
- Spreadsheet exports: keep untrusted labels as text
- Project: review a depot replenishment workbook
- Spreadsheet prompt and workbook release decisions
