The useful outcome was a claim we could actually support.

In 2023 Navirelan held 7.8 million registered mobile-subscription records in the population under review. A presentation described that number as people reached by the service. Farah Mervanti challenged the wording because the records represented subscriptions, and because registration alone did not establish recent use. The communications sponsor wanted a single large number that could travel easily between presentations.

I chose to retain the registered population while adding a separate internal activity view. We defined qualifying activity within a ninety-day window for this analysis and found 6.1 million records meeting that condition, with 1.7 million outside it. The definition and observation period travelled with both figures. This was an internal account, not a claim to reproduce an official statistical indicator or satisfy a reporting rule through a chosen window.

S$3.8 million supported the data work. The team documented which records belonged to the review, how activity was evaluated and where the two populations differed. Commercial colleagues could use the registered figure for a question about the account base and the activity figure for a question about that particular condition. Neither was relabelled as unique individuals. We did not attempt to infer how many subscriptions each person held from the difference between the two totals.

The outcome was a population statement with an explicit definition, rather than a larger reach claim. We gave up the simplicity of one unqualified headline. The 1.7 million records outside the window were not automatically described as lost customers, fraud or records for deletion. Their interpretation required the relevant commercial and operating evidence. I would establish the language of the intended claim before preparing the chart next time. Farah’s challenge arrived after the headline had acquired internal momentum; attaching the counting object and activity condition at the beginning would have made the eventual correction less contentious.

The 2007 usage-data delivery log contained eighty million received records. I approved a management account describing them as traffic events, relying on the apparent completeness of the ingestion log. Daniel Varneshu questioned why a delivery-retry episode coincided with a rise in the count. His review identified seventy-four million distinct source-and-event identifier pairs and six million documented replayed receipts.

My mistake was to treat successful receipt as evidence of an additional underlying event. The delivery team had done useful work recovering interrupted transfers, but that recovery did not create new usage. A proposed quick response was to remove any rows with matching amounts and times. I rejected that after Daniel showed why similar attributes were insufficient: separate legitimate events could resemble one another, while a replay needed evidence connecting it to the same source event.

We used S$720000 for review and correction. The receipt history remained available as a record of what arrived and when. A distinct-event view used the documented source-and-event pairs for the reviewed population. We kept the replay explanation alongside the reconciliation, so a reader could understand why eighty million receipts became seventy-four million distinct pairs without imagining that six million real events had simply vanished.

The corrected account did not establish an invoice error, a refund or an increase in network traffic. Billing owners would need their own evidence for any billing conclusion. Nor did it authorise an unrestricted rule for every future extract: the key and replay evidence had to remain valid in its source context. I accepted responsibility for the initial label and explained the correction to the management recipients. I would now ask explicitly whether a count represents transmission attempts, receipts or underlying events before signing off. The enterprise gained a clearer distinction between delivery reliability and the event account; the difficult consequence for me was admitting that a technically complete log had supported a conclusion I had not earned.

A 2015 customer-history review covered eighty-four thousand number disconnections. Fifty-one thousand numbers were reassigned during the review year. An analytical join used the telephone number as though it identified the same customer throughout the period, bringing records from an earlier assignment into a later one. The field was convenient and widely available, which made the mistake easy to repeat across apparently consistent extracts.

The analytical sponsor preferred a continuous history because it made longer comparisons possible. I asked for the assignment evidence before accepting that continuity. A telephone number could identify the service within an interval without establishing that two intervals belonged to one person. We gave up the convenience of a single uninterrupted number-based account where the source did not justify it.

S$2.4 million funded the data response. The team retained assignment intervals and connected records within the supported relationship, preserving the disconnection and reassignment evidence used to distinguish them. Where timing or identity remained unresolved, the uncertainty stayed visible. We did not claim that every disconnection led to a new person or that every reused number established an entirely unrelated household; the review supported boundaries in the account, not speculation about identities beyond it.

Privacy and legal owners assessed permitted uses of customer information. The data correction did not itself confer permission to combine records, establish anonymisation or settle a person’s rights. No actual subscriber details appear in this fictional example. The delivered change was a more faithful relationship between a number and its dated assignments, with unsupported continuity removed from the analytical view.

I would include reassignment scenarios in the original requirements for any customer-history product. Testing only stable numbers had allowed an easy identifier to stand in for a difficult relationship. The important enterprise change was that the history now required evidence of continuity rather than assuming it from a familiar field. That made some analyses shorter, but made their claim about whose experience they represented more defensible.

The 2021 network inventory review contained twelve hundred physical sites and thirty-six hundred sector records. A summary labelled the sector total as sites, creating the impression of a much larger physical estate. The arithmetic was not the problem. The hierarchy had been flattened when the records moved into a management table, and the label no longer identified the object being counted.

The reporting sponsor wanted to preserve the larger number because earlier presentations had used it. Engineering colleagues asked that the physical-site account remain distinct from its associated sectors. I supported that distinction even though it required changing a familiar series. A sector record could describe an element associated with a site; it did not create another location simply by occupying another row.

We allocated S$1.9 million to this inventory review and retained the relationships in the reviewed inventory. Engineering owners checked the meaning and source of those relationships. The account showed twelve hundred physical sites and thirty-six hundred sector records under their respective definitions. It did not infer a uniform three sectors at every site merely because the aggregate ratio was three. Individual relationships required their own evidence.

The result corrected the inventory statement without claiming construction, coverage, capacity or an improvement in uptime. Those questions depended on engineering and operating evidence that this count did not supply. We also retained the explanation of the earlier label so recipients could understand the series change instead of interpreting it as an abrupt contraction of the estate.

The trade-off was a more demanding presentation: readers needed the two objects and their relationship rather than one large headline. I would involve the engineering owner when specifying the management table, before the hierarchy is lost. A neat table is not a neutral compression when it removes the distinction that gives the number meaning. The corrected account helped leaders ask the next network question without first having to undo an exaggerated picture of the physical footprint.

A 2025 experience review collected twelve thousand voluntary tests, including nine thousand urban and three thousand rural observations. A proposal would have used the results to describe experience across the subscriber population. Participation depended on the devices, locations, times and reasons that led people to run a test. The size of the collection did not remove those selection conditions.

The commercial sponsor wanted an easily communicated all-subscriber statement. The analysis group could explain the observations it held but lacked a basis for treating them as a representative sample of every customer’s experience. I chose to publish an internal cohort account and withhold the broader claim. This did not mean the tests were worthless; it meant their usefulness depended on keeping the observed population attached to the findings.

S$3.1 million supported evaluation work. We described the voluntary collection, the urban and rural counts and the relevant observation conditions. We separated questions the cohort could inform from those requiring a different study. The team did not invent population weights, assume that the urban-to-rural ratio matched subscribers or produce a representative speed result from information that had not established one.

The delivered outcome was a qualified account for further investigation. It supplied neither an overall customer-experience score nor evidence that network performance had improved. Network operations retained responsibility for interventions and would need additional evidence to assess their effects. The cost of the decision was losing a prominent general claim that the sponsor had hoped to use immediately.

I would agree the intended population before deciding how observations will be collected. Starting with a convenient dataset and only later asking it to represent everyone creates pressure to make the evidence fit the announcement. A useful test programme can begin with selected observations, but its conclusions should remain within their reach. Keeping that boundary explicit allowed the review to suggest where further enquiry was needed without presenting those suggestions as settled knowledge about the whole service.