Keep the data property separate from the scientific conclusion it does not establish.

The 2023 portfolio proposal at the fictional Seravune Therapeutics brought 240000 research readings from twelve studies toward a combined view. Nine studies used one measurement method and three used another. Marcus Selvarne wanted a pooled chart that would make the portfolio easier to discuss. The results were expressed in the same reported unit, and the large collection appeared to offer a more comprehensive account than separate study views. I agreed that the portfolio needed a useful presentation. I did not think the shared unit answered the question of scientific comparability.

The measurement method remained part of how an observation had been produced. A consistent column label could make the values easy to combine without establishing that they were suitable for the proposed interpretation. My role was to preserve that context and make the unresolved question visible. Scientists and statisticians needed to assess comparability for the intended analysis. The data office should neither declare the methods equivalent nor conclude that they could never be compared merely because their names differed. Both conclusions would exceed the decision I was qualified to make.

We retained the method families in separate views pending the scientific and statistical review, with S$3.8 million programme expense. The underlying study associations remained attached to the readings. Marcus had to accept a less compact portfolio presentation while the question was examined. I supported that cost because a pooled display could create an impression of comparability before the relevant assessment had occurred. The account did not claim a therapeutic effect, a patient benefit or a product decision. It concerned the organisation and interpretation of research material for a particular proposed view.

The outcome was a qualified portfolio account with the method distinction preserved, not an eventual finding of equivalence or a completed pooled scientific conclusion. The story reports no later resolution that the record does not contain. I learned to ask whether the convenience of a common data structure is being mistaken for agreement about the observations it holds. A CDO can improve access and organisation while still requiring the responsible experts to judge meaning. The distinction is especially important when a visually coherent chart may lead a management audience to believe that a scientific question has already been settled.

A 2021 research-registry review covered 18000 records for a baseline field. Eleven thousand two hundred contained a recorded value; 6800 were missing that value in the material under review. The proposed analytical preparation would have coded the missing entries as normal to create a complete column for an internal comparison. The argument was that an explicit value would make the dataset easier to use and avoid losing a large portion of the records. I objected because absence in the available material did not establish the clinical state the proposed code would assert.

The missingness could have several origins: the observation might not have been collected, might not have reached the source or might not be accessible in the reviewed extract. The data office needed to investigate the available record rather than choose a reassuring clinical interpretation. I did not claim that every missing entry had the same cause or that the source review would recover them all. The question was whether the preparation step would preserve uncertainty or transform it into a finding without supporting evidence. That was a decision about the integrity of the data account.

We rejected the normal-value coding and supported S$2.2 million source and fitness review. The 11200 recorded values and 6800 missing entries remained distinct. Scientists, statisticians and the relevant clinical owners retained responsibility for any analysis or interpretation. Consent, privacy and permitted use also remained with the appropriate processes. The choice made the proposed comparison harder to complete and left a substantial unresolved portion of the registry visible. I considered that preferable to a superficially complete dataset whose added values represented an assumption the record did not justify.

The reported outcome is the retained missingness and the review decision. It is not a clinical conclusion about those 6800 records, a claim that the missing values were later recovered or evidence that any treatment result changed. I would bring this case into a discussion of data leadership because the pressure to make a table usable can change its meaning before a formal analysis begins. A CDO's contribution includes noticing that change and asking the responsible experts how the uncertainty should be handled. Completeness of a column is not the same thing as knowledge of the observation it is supposed to contain.

In 2011, a research-data account contained 2700 laboratory readings from 900 source samples, with three replicate readings for each sample. I described the material as 2700 independent samples. Lina Mervathi asked what the additional rows represented and challenged my independent-sample wording. The laboratory had produced repeated readings, not another 1800 source samples. My account had allowed the shape of the dataset to change the meaning of the underlying material. The arithmetic of three readings per sample was straightforward; the interpretation I attached to the total was wrong.

I initially focused on the value of the repeated measurements. They could be useful, but that did not justify describing them as separate independent samples. Lina's challenge required me to distinguish the number of recorded observations from the sampling unit and leave the relevant inference to the statisticians. I could not repair the claim simply by replacing independent samples with observations in one chart while allowing earlier discussions to retain the larger interpretation. The correction needed to explain both counts and why the original wording exceeded what the data structure established.

We supported S$480000 review and withdrew the 2700-independent-sample description. The revised account retained 900 source samples and 2700 readings, with the replicate relationship explicit. It did not announce that all 900 source samples were statistically independent for every possible analysis; that assessment remained with the responsible specialists. Nor did it imply that the laboratory work was wasted. The useful material was preserved while the interpretation was corrected. My responsibility was to make the record intelligible and own the inaccurate statement I had made about it.

The consequence included revisiting the account with colleagues who had already heard the first version. That was less comfortable than quietly improving a later report, but it was necessary for the correction to reach the claim's actual users. No scientific benefit, changed treatment conclusion or later study success is attached to the review expense. The lesson remained specific: another row can represent another reading of the same source, and the public or management description must not silently change that relationship. I now ask what the observation unit is before allowing a dataset total to become a statement about the strength of research evidence.

In 2017, research colleagues were struggling to locate and understand datasets already held within Seravune Therapeutics. One proposal was to wait for a full platform replacement before investing further effort in the existing estate. That approach promised a more consistent future environment and avoided building an intermediate solution. I saw a nearer-term problem that the platform decision did not settle: many datasets lacked a usable account of their purpose, ownership and source context. Moving them to a new system would not automatically create that description.

We selected a S$1.6 million programme to curate metadata for thirty-seven research datasets, working with research and technology colleagues. The aim was to make the relevant material easier to find and assess through the existing approved processes. It was not a declaration that every dataset was suitable for a new scientific use. Owners still needed to review the proposed purpose and any relevant access conditions. The programme required time from colleagues who were also supporting the wider platform discussion, so the choice had an opportunity cost rather than being a free preliminary improvement.

For a six-month comparison, the review matched twenty-six pairs of internal retrieval requests by dataset and request scope. Median retrieval time changed from eleven business days in the earlier observations to four in the later observations. I reported that bounded result and the shared contribution of research, data and technology teams. The comparison did not isolate metadata as the sole cause, and it did not show that all research work accelerated. It certainly did not establish a patient benefit or a scientific outcome. The measure concerned retrieval for the matched requests.

The decision allowed useful work to proceed without claiming that the platform question had disappeared. It also gave the institution a clearer account of what information about a dataset was needed regardless of where it was stored. Colleagues who favoured waiting were right that an intermediate programme could create maintenance effort. We accepted that burden for the selected scope and retained the larger technology decision separately. The outcome was a practical service improvement with explicit limits. I would use the case to explain how a data leader can choose a useful step between inaction and a comprehensive replacement, while keeping the measured result proportional to the work actually reviewed.

The 2025 archive exercise attempted to replay sixty-four research analyses using their stored code and data. Forty-nine ran again from the retained material. Fifteen could not be replayed because a required dependency version was missing. The initial programme account favoured describing the archive as reproducible, with the unresolved analyses treated as a technical clean-up item. I wanted the 49/15 split to remain explicit. I also wanted the account to avoid implying that successful replay independently validated the scientific conclusion of the analyses that ran.

Running a stored process again can be an important property of an archive. It helps another authorised reader inspect what was done. It does not establish that the original analytical choices were appropriate, that the source observations were free of error or that the scientific interpretation was correct. Those questions require the responsible scientific and statistical judgement. My team could describe replay status and the missing dependencies. It could not convert an operational test into a general endorsement of the research simply because the outputs could be produced again.

We retained the forty-nine successfully replayed analyses and fifteen unresolved cases as separate states, and funded S$2.9 million remedial work. The next programme needed to address the missing dependency versions through the appropriate review, while preserving an account of the original environment. The narrative does not claim that all fifteen were later resolved. Nor does it offer a universal software-retention policy. Technology, scientific and other relevant owners remained responsible for their parts of the work. The CDO contribution was to keep the archive's demonstrated capability and unfinished work accurately described.

The choice gave up a broad completion statement and required continued attention to a smaller unresolved set. It also made the successful forty-nine less open to overinterpretation. I considered both qualifications useful because the archive would otherwise appear to settle more than the exercise had tested. The lesson is that reproducibility of a stored process and independent confidence in a conclusion are related but different questions. A leadership account should state which one the programme examined. The public case reports a specific replay result and a funded next step, without borrowing scientific authority from a technical success or inventing a completed ending.