A polished spreadsheet can hide weak evidence. Two cells with the same number may have very different foundations: one is backed by an invoice, the other is someone’s recollection. An evidence register makes that difference visible without pretending every input can be known precisely before a decision.
Give each input a status
Use four plain labels: measured, quoted, assumed and unresolved. Measured values come from a defined observation or record. Quoted values come from a current provider or supplier offer with scope and conditions. Assumed values are planning choices. Unresolved means the model cannot yet support a number. Do not turn unresolved into zero just because a formula needs a value.
For every input, record its unit, time period, source date, scope and owner. A postage invoice per envelope cannot be pasted into a per-employee field. A quote for another legal entity or service level may not apply. A six-month time study can become stale after a system change. These are evidence mismatches, not minor spreadsheet formatting issues.
A compact fictional register
- Paper distributions: 2,600 per year; measured from 26 completed runs; excludes off-cycle payments.
- Material cost: $1.10 per check; assumed for the lesson; replace with current invoices.
- Handling time: 3 minutes per check; assumed; verify with an observed sample of full runs.
- New annual fee: unresolved; obtain a written quote with its billing unit and minimums.
- Participation: 60 percent of current recipients; scenario only; not a commitment or requirement.
That register immediately reveals that the example cannot establish real savings. Its purpose is to identify what must be learned. A dashboard that reports a precise return on investment from these inputs should be marked as a scenario, not a verified business result.
Preserve disconfirming evidence
If the new workflow takes longer during the pilot, retain the observation and investigate the reason. Do not discard an unfavorable run merely because it differs from the expected benefit. At the same time, separate documented temporary learning work from ongoing effort. Write the classification rule before looking at which version improves the outcome.
Use a short change log when inputs are updated. Record the old value, the new value, the reason and the effect on the conclusion. This prevents an apparent improvement from being attributed to the program when the model actually changed its wage rate, payment volume or cost boundary.
Define a decision and a review point
Before launch, state which unknowns are material enough to delay approval. After launch, compare the predicted quantities with observed quantities over a specified period. Measure distribution counts, actual charges and hours using the same definitions as the baseline. A difference in cost without a consistent denominator cannot validate the prediction.
Corpay’s public PayCard claims may help identify questions, but the employer’s records and applicable contractual terms must answer them. This publication offers an original evaluation method, not a provider ranking or endorsement. A defensible recommendation includes what is known, what remains assumed and what finding would cause the team to reconsider.
A reviewer can test the model backward
Start from the headline result and trace each total to its inputs. For the fictional $6,760 baseline used elsewhere in this publication, the reviewer should find 2,600 distributions, $1.10 in direct outlay per distribution, 3 labor-minutes per distribution and a $30 hourly planning rate. The calculation should reconcile without hidden constants. If a total cannot be reconstructed, the evidence register is incomplete even when every cell has a source note.
Then test each source against its use. A supplier invoice may substantiate materials purchased but not materials consumed in a particular run. A time sheet may establish paid hours but not the portion spent on distribution. A provider quote may cover one program configuration but not another. Source quality is not a simple ranking from invoice to estimate; it depends on whether the source actually measures the input the model needs.
Define a review threshold in advance
A team might decide that an unresolved recurring charge capable of changing the result by $500 is material, while a small stationery uncertainty is not. That threshold is a project-specific judgment, not a universal rule. Record who selected it and why. This helps focus research effort on the assumptions that can change the recommendation instead of spending equal time on every minor input.
The register should also identify the person responsible for an update. Payroll may own distribution counts, procurement may own the quote, finance may own the valuation method and an appropriate legal reviewer may own requirements that constrain the options. Responsibility for supplying a fact is different from authority to approve the entire proposal. Keeping those roles distinct prevents a single completed cell from being mistaken for broad sign-off.
Archive a dated decision version before changing assumptions for the next review. Preserve the original forecast alongside observed results, then explain differences in volume, price, timing and process. Replacing the forecast with actual numbers makes the final model look accurate but destroys the ability to learn whether the decision process was sound. A useful evidence register supports that learning rather than merely documenting a favorable answer.