A source-based spot-check compares selected reported values with relevant supporting records and traces any discrepancy. Define the population, fields, selection method and evidence reviewed. Report observed findings without treating a small purposive sample as proof that every unreviewed record is accurate.
Specify what accuracy means for this check
Accuracy concerns agreement with the relevant underlying fact, not merely a valid format. Choose the fields and reference evidence required for the decision: an invoice amount, a service completion date or a customer assignment. A source export is not always independent evidence if it contains the same transformation being tested. Identify where the original fact is captured and what limitations remain in that record.
Define the population and cut-off before selecting cases. State whether the check covers transactions issued during a period, records received by a date or currently open cases. Include the version of the report and source extract. If the source can change after selection, preserve an authorised snapshot or evidence reference so a later reviewer can distinguish a new correction from the value originally examined.
Choose cases for a stated purpose
A purposive spot-check can deliberately include refunds, large transactions, unusual dates and different source systems to explore likely failure modes. A random sample supports different inferences when its design and size are appropriate. Do not describe a targeted sample as statistically representative simply because cases came from several categories. Record both the selected cases and why they were selected, including any unavailable supporting evidence.
Use a comparison sheet that lists expected value, reported value, reference record and outcome. Distinguish confirmed agreement, confirmed discrepancy and unresolved evidence. A missing document is not a pass. When one mismatch appears, trace its transformation path and test whether a common rule affects other records. The investigation can then expand along a demonstrated mechanism rather than applying an arbitrary multiplier to the initial sample error count.
Report findings at the scope actually reviewed
Describe how many records and fields were checked, how many discrepancies were observed and what follow-up occurred. If two of twenty deliberately selected difficult cases fail, that is an observed sample result; it is not automatically a 10% population error estimate. Estimating an error rate requires a suitable sampling design and treatment of uncertainty. Avoid a headline that implies audit assurance the exercise was not designed to provide.
The linked framework provides accuracy background, while the example is fictional. Keep corrections reversible and distinguish repaired sample records from a verified system-wide fix. After repairing a transformation, rerun affected calculations and a focused source comparison. Name residual gaps, including records without independent evidence, so the publication decision reflects what was checked rather than the comfort of having completed a checklist.
For disagreements with the reference evidence, preserve the explanation of why that evidence is considered authoritative. Two disagreeing exports from the same faulty mapping do not provide independent confirmation.
Fictional transaction review: one date rule is wrong
A fictional scenario to explain the method, not a customer result or a market benchmark.
A fictional dataset contains 10,000 transactions. An analyst selects 20 cases purposively: ten ordinary purchases, five refunds and five unusually large orders. Amounts agree with supporting records in all 20, but three dates from one export are reversed under an incorrect day/month convention. The report records the observed field-level discrepancy and its shared source.
The team identifies 600 records processed through that mapping and separates ambiguous dates from dates whose interpretation can be established. It fixes the parser, rebuilds the affected population and checks additional boundary examples against source evidence. It does not assume all 600 were wrong merely because they used the same parser.
The final note states that 20 initial cases were checked and explains the targeted follow-up. It does not claim all 10,000 transactions were audited or extrapolate a population error percentage from a purposive sample. The invented case illustrates proportionate investigation and appropriately limited claims.
TAKE THIS INTO YOUR NEXT CONVERSATION
A practical next-step checklist.
- Define population, fields and reference evidence.
- Record selection method and unavailable evidence.
- Trace discrepancies to common mechanisms.
- Limit claims to the reviewed scope.
Common questions.
Can twenty checked records establish the error rate?
Not without an appropriate sampling design and uncertainty assessment. A purposive review is useful for finding mechanisms, not automatic extrapolation.
Does a valid date format establish accuracy?
No. A correctly formatted date can still refer to the wrong day, event or transaction.
Further reading & source context.
These primary references provide background. They are not evidence for the fictional example above and do not endorse Accuralytix.
Found an error or a missing qualification? Send an editorial correction.
Check my data readiness.
Use a practical tool to turn this idea into a clearer starting point.
Check my data readiness