On this page

The short answer

Retrieve the complete meaning of a number: value, unit, period, entity, and qualifying condition. Check any calculation separately so correct arithmetic on the wrong evidence cannot pass as a correct answer.

What to take away

  • Keep values attached to labels.
  • Show operands for calculations.
  • Distinguish missing from zero.

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

Define a numerical evidence bundle

Microsoft’s layout reference describes structured document extraction. For numerical retrieval, the test should establish whether values remain linked to their labels and scope. Store the source amount, scale, currency or unit, reporting period, entity, and applicable footnote. A vector hit containing the requested digits is not necessarily the right fact. Decide whether the application may convert units or perform arithmetic and record the assumptions needed for those operations before scoring an answer.

Evidence: Document Layout skill [1]

Create a scale-and-period trap

Author a fictional report listing first-quarter usage as 12 thousand units and second-quarter usage as 15 thousand units. Ask for the increase in individual units. The expected arithmetic is 3 thousand, or 3,000 units, with both periods correctly identified. Add a similarly named business unit reporting monthly figures. The fixture fails if the answer subtracts different periods or mixes entities, even when the resulting subtraction is mathematically correct.

Evaluate lookup before computation

First check whether retrieval supplies the correct operands and their labels. Then inspect normalization, conversion, and arithmetic as separate steps. If the application uses a calculation tool, preserve its inputs and output; if it performs no verified calculation, do not describe the answer as tool-validated. Compare rounding rules explicitly, especially when source values are already rounded. A precise-looking answer with many decimals can imply accuracy that the retrieved evidence does not support.

Include missing and incompatible evidence

Test blank values, dashes, estimates, ranges, and a number marked provisional. Add a case where one period uses a changed measurement definition. The expected behavior may be to qualify the comparison or decline to calculate a single figure. Do not let a judge infer that an absent value is zero. Mark whether the question can be answered exactly, approximately, or not from the eligible evidence. This makes honest uncertainty part of the numerical acceptance contract.

Keep an inspectable calculation record

For critical answers, retain the source locations, operands, units, formula, rounding rule, and final wording. A reviewer should be able to reproduce the reasoning without searching the entire document corpus. Group failures into wrong evidence, wrong unit, wrong period, wrong operation, and overstated precision. Repair the corresponding stage, then replay both the failed fixture and nearby valid cases. A generic answer-similarity score is a weak substitute for this structured numerical evidence trail. Include a percentage change and an absolute difference as separate requested operations. For the same operands, both answers can be numerically valid yet only one satisfies the question. Preserve the operation label alongside the computed result.

Limits of the evidence

The figures are authored arithmetic examples, not financial or operational results. The method does not certify source accuracy or authorize calculations outside the application’s scope.

Common questions

Can exact answer matching validate numerical reasoning?

It can catch differences, but also inspect operands and units because a matching number can arise from wrong evidence.

Should every missing value become zero?

No. Preserve missingness unless the source explicitly defines its meaning.

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 →