The choices hidden inside apparently simple data products

In 2019 I approved a dashboard transformation that selected the highest result from a repeated test series. Commercial colleagues wanted a single value they could explain quickly, and I treated the selection as a presentation choice. Imran Teqvermi challenged that assumption. A repeat might follow a documented procedural issue, a client request or another circumstance requiring professional interpretation. The software could not decide that the largest observation was the right account merely because it produced a convenient summary. I had allowed a data-product decision to erase evidence that belonged in the technical discussion.

We reviewed ninety fictional series. Sixty-one contained a single observation set and twenty-nine involved repeats. Of the repeated series, twenty-one had a documented reason and eight needed further examination. Those categories did not identify twenty-one automatically valid replacements or eight failed materials. They showed where the explanation could be inspected and where the qualified owner still had work to do. Imran wanted the complete sequence retained. Sales colleagues feared that several values and qualifications would make the product seem indecisive, especially when a client wanted an immediate comparison.

I withdrew the highest-result rule and accepted ₹28 lakh of correction work. The revised product retained the series, the reason for repetition where recorded and the status assigned through the laboratory's authorised process. Unresolved explanations remained visible for review instead of being filled with a convenient assumption. We did not delete the initial observations or instruct the system to prefer the lowest value as an equally arbitrary substitute. The technical owner determined whether an observation could be used and how any authorised conclusion should be reported.

The consequence was a less compact screen and additional work for the product and laboratory teams. It did not produce a claimed improvement in material quality or establish that every earlier report was wrong. I had to explain why the initial convenience was no longer defensible and accept the cost inside my data programme. The case changed the approval question I ask: does a transformation only represent an existing decision, or has it quietly made a new one? A data leader needs to recognise that distinction before a polished interface gives an unsupported selection the appearance of professional authority.

In 2022 an evidence export required a numeric value in every row. The proposed mapping converted a below-reporting-limit qualifier into zero so that the receiving spreadsheet could calculate a complete average. In the selected fictional set, forty-six of sixty-four results had quantified values and eighteen carried the below-limit qualifier. The proposed export would have made those eighteen appear to be measurements of absence. It would also have separated them from the limit and method context needed to interpret what the laboratory had actually reported.

The account team wanted a file that would work immediately with the client's existing analysis. They worried that text qualifiers or separate fields would be treated as missing data and make the service appear less complete. Imran's concern was different: replacing a qualification with zero changed the meaning of the observation. The appropriate treatment in an analysis depends on the question, method and available evidence. Neither the data team nor the client-facing export should silently choose a universal replacement value merely to satisfy a numeric column.

I approved ₹31 lakh to revise the exchange format and ₹6 lakh annual support for its interpretation and maintenance. The revised file carried the reported value where available, the qualifying status and the associated limit and method reference. We worked through the receiving system's requirements with the responsible analyst instead of claiming that a more detailed file automatically solved the statistical question. The eighteen qualified results remained part of the record; they were neither discarded nor presented as eighteen measured zeros.

We gave up the promise of a drop-in export that required no conversation about meaning. Some downstream calculations needed revision, and that work was a genuine cost to the service. The result was a more faithful exchange, not a finding that any actual material was safe, unsafe or free of a constituent. The numbers describe an invented selection, with no environmental or health assessment implied. For me, the leadership question was whether we would preserve a difficult qualification when a simpler representation made the product easier to sell. A complete column is not a sufficient reason to manufacture precision the observation never supplied.

In 2023 a proposed predictive product used 240 fictional specimens from forty material batches, with six specimens per batch. The initial split allocated 180 specimens to development and sixty to assessment. Thirty batches appeared on both sides, with four specimens in development and two in assessment; the remaining ten batches contributed all six specimens to development. The split was arithmetically correct, but the intended commercial use concerned a new batch. A model could therefore be assessed on related observations from batches it had already encountered, answering an easier or simply different question.

The analytical team had invested time in the original workflow. Commercial colleagues wanted to retain the assessment already prepared for a product discussion, arguing that the held-out specimens were still separate physical observations. That was true, but it did not establish that the assessment represented the future decision. Imran asked whether the proposed use could tolerate the difference between a new specimen from familiar conditions and a batch whose shared conditions were absent from development. I supported the need to make the validation unit explicit before promoting the result.

We reassigned complete batches: twenty-six batches, or 156 specimens, for development and fourteen batches, or eighty-four specimens, held apart. The ₹45 lakh validation programme and ₹7 lakh annual review provision funded the redesign and subsequent scrutiny. The revised allocation did not by itself establish model quality, and we did not invent a higher score to justify the delay. Preprocessing and model choices also needed to respect the separation. The qualified analytical team retained responsibility for the eventual assessment and the limits of any approved application.

The commercial consequence was that the earlier assessment could no longer support the proposed launch account. We accepted further work without knowing whether the later result would strengthen or weaken the business case. That is the trade-off I want a board to understand: improving the relevance of evidence can remove confidence rather than add it. The case is not an instruction that every model needs this exact split. It shows why the unit of independence should follow the intended use, with related observations treated honestly, before a convenient metric becomes a product promise.

In 2024 a client portal proposal displayed rounded measurements beside a coloured conformity category. The product team intended the display to be easier to scan than a technical report. In a fictional review of twenty-four proposed result displays, seventeen reproduced an already authorised category while seven needed the laboratory's decision rule and supporting context before a category could be assigned. The software's rounding and comparison logic had begun to supply that category itself. A measured value, its uncertainty and an authorised conformity statement were being collapsed into one interface decision.

The product lead wanted consistent colours and a predictable response for every row. Leaving some entries without a simple verdict would require another explanation and could look unfinished. Imran argued that the appearance of consistency was not a sufficient basis for changing a technical decision. The applicable method, specification and agreed reporting process needed their own assessment. A rounded number on a screen should not become an alternative route to a conclusion that the qualified owner had not made.

I committed ₹26 lakh to revise the portal contract between the data layer and the authorised reporting process. Where a category had been approved, the interface could display it with the appropriate context. Where the decision remained unresolved, it retained that state and directed the question to the responsible owner. The data team did not invent a universal guard band, choose a tolerance or claim that more decimal places would resolve every uncertainty. We kept the underlying record and the presentation choice separately inspectable.

The change cost us a uniformly coloured prototype and delayed the simplified comparison feature. It did not identify seven failed materials or retrospectively certify the seventeen other displays. Those counts describe the review state of invented examples, not a quality rate. I learned to ask whether an interface has authority to make the conclusion it encourages its reader to draw. A useful summary may still be concise, but its decision must come from a process entitled to make it. Product elegance becomes fragile when it depends on concealing who actually owns the judgement.

In 2025 the commercial team proposed a reusable evidence product drawn from forty-eight fictional client projects. The archive was already in our systems, and the product outline treated that possession as a starting point for resale. A project-by-project review found twenty-nine with a documented basis for the specified reuse and nineteen with unresolved permissions or scope questions. Removing client names would not, by itself, answer those contractual questions. The proposed archive also reflected the projects the practice happened to receive, rather than a designed sample of an entire market.

The commercial sponsor wanted to retain all forty-eight projects because a larger dataset would be easier to describe as a benchmark. The nineteen exclusions reduced coverage and made the initial offer less attractive. Our contract and legal owners wanted the intended purpose, restrictions and approvals assessed before the new product was offered. Imran also challenged the implied representativeness. Even a fully authorised collection could still be selective in material types, project circumstances and the reasons testing had been requested.

I supported ₹34 lakh of mapping and review, with a ₹9 lakh annual stewardship provision for the bounded product. The initial release scope used only the twenty-nine projects cleared for that specified purpose, subject to their recorded conditions. The nineteen unresolved projects remained outside it pending an appropriate decision. My team linked the inclusion record to those permissions and retained a description of the collection's limits. We did not substitute a data-team opinion for legal review or call the resulting subset a representative industry measure.

The consequence was a smaller commercial proposition and additional continuing work. No revenue uplift or customer adoption is invented to make the decision appear painless. We gave up some of the breadth that had made the original pitch attractive. The case asks whether an enterprise can distinguish an archive it holds from a product it is entitled and equipped to offer. I want directors to examine that distinction when data is described as an asset, including the permissions, selection and maintenance work that determine what the asset can responsibly become.