AI percentage rounding audit: explain a displayed 101%

Define a complete single-choice set before rounding

AI means artificial intelligence. An AI percentage rounding audit starts with source counts, not with a desire to make the displayed labels add to a tidy number. In this exercise, seven fictional delivery-preference cards each contain exactly one category. Every supplied card belongs in the calculation. No real people were surveyed, and no company report or AI system was tested.

Write the contract: “Practice V1; eligible IDs B1–B7; one card is one unit; categories D, E and F; one category per card; all cards included; common denominator seven; display each percentage rounded independently to the nearest whole percent.” A denominator is the count you divide by. A mutually exclusive set means no card belongs to two categories; exhaustive means every eligible card has a category.

The source counts stay exact throughout. Rounding applies only to the calculated percentage labels. This is not a privacy procedure that rounds the counts themselves. We also avoid halfway cases: none of the percentages in this card ends exactly halfway between two whole numbers. Different software rules for ties therefore do not affect the exercise.

The business data analysis guide explains the broader habit of defining records and checking calculations. Your smaller deliverable here is a saved source-to-display ledger. It will distinguish an exact total of 100% from a displayed label sum of 101%, without changing any card.

Preserve the seven source IDs and their categories

Treat D, E and F as three invented delivery-preference labels. They have no ordering and do not stand for good or bad service. These cards are deliberately constructed teaching inputs, not a random sample or an estimate of demand. The only question is what the stated source and display rule support.

Check coverage before calculating a percentage. List the distinct IDs: B1, B2, B3, B4, B5, B6 and B7. There are seven. Then group the IDs: D contains B1/B4, E contains B2/B5, and F contains B3/B6/B7. Counts are two, two and three. Their sum is seven, matching the original card count.

Check membership as well as the summed counts. A total of seven would not, by itself, rule out counting one ID twice while omitting another. Here every original ID appears once in the grouped lists. Keep those lists beside the counts so another reader can verify the completeness claim rather than simply trust the heading.

Do not replace the source card with a percentage-only summary. If you later discover a real coverage problem, correct the source under a documented new version and recalculate. The spreadsheet prompt examples can help describe such a workflow, but the present audit makes no changes to a live workbook.

Calculate exact fractions before choosing display labels

Use a calculator to evaluate 2 divided by 7 multiplied by 100. The decimal begins 28.571428…%. Its exact expression is 200/7%. Category E has the same count and the same fraction. For F, calculate 3 divided by 7 multiplied by 100: 42.857142…%, exactly 300/7%. Preserve these fractions instead of treating a shortened decimal as the source value.

Add the exact shares first: 2/7+2/7+3/7=7/7=1, or 100%. Equivalently, the exact percentage sum is (200+200+300)/7=100%. Now add only the displayed whole-number labels: 29+29+43=101%. These are two different calculations. The displayed sum is one percentage point above the exact total.

Record both results with names: “Exact source-share total: 100%. Independently rounded label total: 101%.” The extra displayed percentage point is not an eighth card. It arises because rounding each category and then adding is not the same calculation as adding the exact category shares first.

This demonstration does not imply every non-100 total is harmless. Its explanation works because the source IDs, common denominator, category coverage and stated display rule have all been checked.

Separate supported explanations from convenient guesses

Claim one says: “A total of 101% proves a duplicate card.” Reject that conclusion for this supplied set. All seven IDs are unique, each appears once, and the exact fractions total 100%. The displayed 101% is reproduced by the independent nearest-whole rounding rule. Another dataset could contain duplicates; this result is not permission to skip its ID check.

Claim two says: “Change F to 42% so the labels total 100%.” Under the stipulated rule, 300/7% rounds to 43%, not 42%. Silently changing that label would no longer follow independent nearest-whole rounding. Keep the supported labels and add a clear note explaining the rounding difference instead of presenting a different rule under the same heading.

There are separate presentation methods designed to make displayed parts sum to 100. This exercise neither implements them nor declares all of them wrong. If such a method is selected, describe and check it explicitly. The unacceptable step here is quietly changing the source or the method merely to hide a visible discrepancy.

Claim three says: “Any percentage mismatch is only rounding.” That is also unsupported. Missing IDs, overlapping categories, multiple-choice answers and inconsistent denominators can produce different problems. Record those as checks, not as findings about our fictional card. The resolution cohort audit illustrates how inclusion and outcome definitions change what a rate means.

Use official rounding explanations without importing their policies

Ofqual’s published rounding policy explains that a total calculated from original values may differ from the sum of separately rounded entries. It also describes deriving percentages from unrounded values and notes that the displayed percentages may not add to exactly 100%. That mathematical presentation principle supports keeping an exact source calculation separate from shortened labels.

Ofqual also has rules about rounding counts and statistical disclosure control. Those are not the rules of our seven-card exercise. We are not applying an education-statistics confidentiality policy to business data, stating a legal reporting requirement, or claiming this fictional card represents an official survey. Our counts are unchanged integers and our chosen display precision is a whole percentage.

Microsoft documents ROUND as rounding a number to a specified number of digits, with zero decimal places producing a nearest integer. That is enough to explain the display operation used here. No spreadsheet function was executed for the article, and the manual core does not depend on a product version or subscription.

The earlier chart and percentage claim audit checks relative changes, share differences and chart scales. This exercise is not another chart inspection. Its output reconciles all parts of one complete category set: source IDs, exact fractions, independent rounded labels and their different summed totals. The distinction matters when deciding which checking note to keep.

What can another reader reconstruct from your saved file?

On Windows, open Start, type Notepad and launch the application. Create a document headed “Percentage display audit V1: fictional preference cards.” Include the contract, all seven source IDs, category membership lists, counts, exact fractions, displayed percentages and both named totals. Add a line for each of the three claim decisions, with the source evidence that supports it.

Press Ctrl+S, choose Documents and save the file as percentage-rounding-audit-V1.txt. Close the editor and reopen the document from that folder. Check that B7 remains present, F still has three source IDs, and exact 100% is not relabelled as displayed 101%. Confirm that the unresolved possibilities for other datasets have not been turned into invented problems in this one.

For an optional AI comparison, provide only these fictional cards and the contract to an assistant you already use. Copy this instruction: “List every ID under its supplied single category. Use seven as the common denominator. Show source counts, exact fractions and independently rounded whole-percentage labels separately. Calculate the exact share total and the displayed label sum. Do not change a count or silently adjust a label to force 100%. Assess the three claims against the source.” Compare the response with the manual answer key.

If access is unavailable, write “optional AI comparison not performed.” The manually checked file is still the core result. No survey is launched, no customer is contacted and no production report is edited. A suggested fifteen-minute allocation is three minutes for scope and source, four for membership and counts, four for fractions and claims, and four for saving and reopening. This is a planning allowance, not measured speed.

In my earlier writing about perfection and effectiveness, I described projects delayed by trying to perfect every detail. Applied here, the useful habit is to finish a small explainable record rather than postpone checking until a reporting system is redesigned. This is an application of that working principle, not a claim of a personal percentage-report experiment or a client outcome.

Close with an acceptance note: “Seven cards are fully covered. Exact shares total 100%; independently rounded labels total 101%. Source values and the agreed display rule remain unchanged.” For a different situation where overall and subgroup rates answer different questions, use the segment versus total audit. Do not use its explanation as a substitute for this rounding calculation.