On this page

The short answer

Test spreadsheet retrieval on the meaning of a cell within its workbook, sheet, period, and unit. Preserve whether a value is authored or calculated, and do not assume a stored formula value is current.

What to take away

  • Keep workbook and sheet lineage.
  • Distinguish formulas from values.
  • Check units and missingness explicitly.

Sources checked . Research synthesis and Rubrex recommendations. Examples and local experiments are labeled; they are not client results.

Define what the extracted representation means

Microsoft’s Document Layout skill lists spreadsheet input among supported formats. That does not establish full spreadsheet calculation semantics. Decide whether your pipeline reads displayed text, stored values, formulas, or a structured export. Record workbook revision and sheet name with each evidence unit. A cell containing 15 can mean a percentage, a count, or a value in thousands. Its surrounding labels and formatting may carry meaning that a plain-text dump does not retain.

Evidence: Document Layout skill [1]

Build a sheet-identity challenge

Author a workbook with Actuals and Forecast sheets containing the same regional rows. Put quarterly revenue in thousands on Actuals and monthly projected revenue in units on Forecast. Ask for the actual quarterly north-region amount. The acceptance rule requires the right workbook revision, sheet, period, and unit. Add a formula total whose stored value is deliberately not refreshed in the fixture. The application should not present that stored value as a verified recalculation it never performed.

Separate retrieval from calculation

For a lookup question, verify that the retrieved evidence identifies the correct cell and its labels. For a computed answer, retain source operands, formulas or explicit arithmetic, missing values, and the calculation environment. A language model repeating a displayed total is not evidence that the workbook was recalculated. If the application does not evaluate formulas, state that boundary and test the corresponding qualification. Reject a confident result when the required inputs or calculation semantics are unavailable.

Test structural features that change meaning

Include merged headings, hidden rows, filters, named ranges, and footnotes when they affect the intended task. Do not assume a hidden row is unauthorized or irrelevant; define the application’s eligibility rule separately. Check whether a total already includes a subtotal before summing both. Test a blank cell, a dash, and a genuine zero as distinct values. These cases can produce convincing numerical answers while the underlying evidence has been interpreted incorrectly.

Keep a source-level audit trail

Save the sheet identifier, range, revision, and extracted representation with each failed case. Compare parser changes against those anchors rather than matching answer text alone. Require critical numerical answers to preserve unit and period and to disclose unverified calculations where relevant. If the representation cannot support a reliable cell relationship, route the question to a structured calculation path or an explicit limitation. A general-purpose retrieval index should not silently impersonate a spreadsheet computation engine. Test duplicated sheet names across different workbooks. A source link containing only the sheet name can appear correct while pointing to the wrong workbook revision or business entity.

Limits of the evidence

This is an authored workbook test plan, not a spreadsheet-engine or financial audit. Stored values, formula evaluation, and formatting behavior depend on the extraction path.

Common questions

Can a formula string prove the numeric answer?

No. Verify the inputs and calculation or clearly identify the result as a stored, unverified value.

Should hidden rows always be ignored?

Only according to a defined task and eligibility policy; visibility is not itself an authorization rule.

Sources & further reading

Primary sources behind this briefing. A source’s findings apply to its own study conditions; publication on arXiv does not establish peer review.

  1. Document Layout skill Microsoft · Living reference · Maintained documentationReviewed: structural extraction, supported representations and location metadata. Accessed October 16, 2026.

Questions or corrections? Write to Rubrex. Read our editorial approach.

From reading to a repeatable evaluation

Put a quality bar around your use case.

The AI Reliability Sprint covers one use case, a baseline, failure analysis, and one improvement experiment in 10 business days.

Request an evaluation call

See the Sprint deliverables →