I want the measure to retain the event and undertaking it describes.

The 2023 delivery review covered twenty thousand shipments. Sixteen thousand met their original promises and four thousand were late against those commitments, giving an original-promise on-time result of 80%. Dates had subsequently been revised for three thousand of the originally late shipments. Against the updated promises, nineteen thousand now qualified as on time, or 95%.

Mara Ellervane questioned a presentation showing only the updated-date figure. The service sponsor preferred that account because the revised promise was the one colleagues were currently discussing with customers. I agreed that current commitments mattered to the next action, but declined to let them erase the earlier undertaking. A 15-percentage-point difference between definitions did not establish that the physical delivery had improved by the same amount.

We used US$2.8 million for data work retaining original and revised promises, the change time and the recorded reason. The two performance views remained separately labelled. Commercial and service owners could explain the circumstances of a revision, including what had been agreed, without the data function making a contractual judgement from the date change alone. We did not describe every revision as improper or assume that the original commitment remained the only useful operating reference.

The outcome was an account of both the original undertaking and the current position. It cost the sponsor a single uncomplicated success figure. It did not produce a measured acceleration of the journeys, a customer-retention result or an automatic liability conclusion. I would agree the promise-history design before the next performance series is launched, rather than recover the distinction after a revised-only percentage has become familiar. The enterprise needed current information and an honest history together. Preserving both made it possible to discuss service recovery without rewriting the basis on which the earlier performance had been understood.

In 2001 I reviewed an export containing fifty-two thousand rows and approved it as a shipment total. Caleb Vornesley traced the file structure and found eighteen thousand consignment headers together with thirty-four thousand piece-level rows. The extract contained two levels of a hierarchy. Adding them did not create fifty-two thousand separate consignments, even though every row was a real record in the file.

The reporting team initially proposed removing the piece rows from the export to prevent another mistake. I wanted the management count corrected while preserving the detail needed by operating colleagues. The error lay in the interpretation I had approved, not in the existence of the piece records. Deleting that useful level would make the summary easier at the cost of information the business still needed.

US$260000 supported the correction. We separated the consignment and piece accounts and retained their parent-child relationships. The explanation identified what each count represented and why the earlier total should not be used as a shipment number. We did not infer a uniform number of pieces per consignment from the aggregate ratio. Individual consignments required their own relationships, and other operating volumes could not be reconstructed by treating the row total as an interchangeable measure.

I took the corrected interpretation back to the management recipients of the original figure. The work did not establish additional freight, a revenue adjustment, a refund or a productivity gain. Those conclusions would have required evidence beyond the hierarchy review. The uncomfortable consequence was admitting that my approval had turned a valid export into an invalid headline.

I would now require the object being counted to appear in the review before discussing the total. A populated identifier and a plausible volume are insufficient when an extract combines levels. The lasting change was a distinction between the detail needed to describe a consignment and the number of consignments themselves. Caleb’s challenge improved the account because we preserved that relationship rather than conceal the error by removing its supporting detail.

A 2014 review followed one thousand trailer visits. The median interval from gate entry to gate exit was three hours, and a proposed account labelled it unloading time. Only 620 visits had paired unloading-start and unloading-end observations. The remaining 380 lacked that pair. Gate timestamps described the whole visit, which could include waiting, unloading and other activity; they did not isolate each part.

The operating sponsor wanted a common unloading measure across all thousand visits because it would make comparisons easier. I declined to fill the missing unloading intervals with the available gate duration. That would make the population complete by changing the activity measured. It could also direct attention towards the unloading team for time whose cause the account had not actually established.

US$1.6 million funded the measurement review. We retained the whole-visit interval for the thousand visits and a separately scoped unloading account for the 620 with the required observations. The 380 without paired unloading timestamps remained explicit. We did not invent a median for the observed unloading subset, allocate the gate interval among activities without evidence or infer that the observed subset represented the missing visits.

The output clarified which duration could answer which question. It did not reduce dwell, prove that unloading was slow or establish a new operating standard. The service and depot owners retained responsibility for investigating causes and choosing interventions. The trade-off was that management lost a single comparable-looking measure for the complete population and gained two accounts with different boundaries.

I would include the required activity observations when specifying the review, before assuming that the most available timestamps answer the intended question. A whole visit can be a valuable measure in its own right. The problem was not that gate data were poor, but that their label assigned them a narrower meaning. Keeping the interval honest helped operating colleagues decide what further observation was needed without attributing unmeasured time to the wrong task.

The 2020 review used a simplified two-route freight account. Six thousand US short tons travelled one hundred miles, and four thousand short tons travelled four hundred miles. The correct sum was 600000 plus 1600000, or 2.2 million short-ton-miles. A proposed calculation averaged the two route distances to 250 miles and multiplied by the combined ten thousand tons, producing 2.5 million instead.

The sponsor liked the shorter calculation because the distance average was easy to explain. I asked the team to keep each weight paired with the distance it actually travelled. The routes did not carry equal weight, so their unweighted mean distance could not stand in for the weight-based calculation. The corresponding weight-weighted distance was 220 miles, not 250. The account used US short tons of 2000 pounds, not metric tonnes.

We allocated US$2.2 million to the broader data work establishing the weight-distance basis and its source relationships. For this reviewed example, the calculation retained the two contributions before aggregation. We documented the unit and scope rather than allowing the familiar word ton to conceal a different mass convention. The response corrected the freight measure; it did not change either route, the cargo moved or the physical distance travelled.

The result also supplied no emissions, cost or productivity improvement. Those questions would require additional inputs and suitable methods under their respective responsible owners. A smaller corrected transport-work figure cannot by itself be presented as fuel saved or a more efficient network. The cost of the decision was abandoning an appealing shortcut and explaining why two apparently reasonable summaries disagreed.

I would preserve paired quantities in the initial analytical specification, especially when the final measure depends on their product. Aggregating each side separately can discard the relationship the calculation needs. The example is deliberately simple so that a senior reader can examine the difference directly, without mistaking a measurement correction for an operational achievement.

The 2025 delivery account contained thirty thousand recorded events. Supporting material included twenty-seven thousand signed-proof records, twenty-four hundred third-party scans and six hundred events without supporting evidence in the review. A proposed status field described the complete population as signed delivery proof, using the existence of a recorded delivery event to stand in for the document supporting it.

The service sponsor wanted one closed status that downstream teams could use easily. I chose to preserve the difference between the event and the evidence associated with it. A third-party scan could be useful, but it was not automatically the signed record described by that label. An event lacking reviewed evidence did not establish that delivery had failed; it established the limit of the supporting account available to us.

US$3.6 million supported classification by source and evidence status. We retained the three groups and made the link between each event and its supporting material inspectable within the authorised process. Claims, commercial and legal owners assessed the implications for their decisions. The data team did not infer liability, approve refunds or decide contractual sufficiency simply by changing the label.

The resulting account could support a more precise follow-up. Teams could ask for the evidence actually missing or examine whether the available source met their purpose, instead of assuming that every recorded event came with the same documentation. No improvement in collections, claims cost or physical delivery was asserted. The trade-off was losing a convenient single status and accepting additional explanation for users who had expected the account to be uniformly closed.

I would separate event occurrence and evidence status when designing the record, before the same field is asked to answer both questions. The distinction matters even where most events have signed support. The largest group should not lend its documentary status to every remaining row. A clearer account helps the responsible owners assess the case in front of them without making the data label a substitute for their judgement.