In 2018 I approved a research summary describing the average energy-only price across a combined purchased volume. The underlying fictional contracts covered one MWh at ₹4,000 per MWh and nine MWh at ₹6,000. We copied both rates accurately, then used their equal-contract mean of ₹5,000. The label implied a pooled-volume result. My approval allowed an exact calculation to support a claim that the calculation did not answer. The contracts shared the same study period and price basis; this was not a comparison of unrelated bills.

Nitin challenged me to follow the quantities as well as the rates. One contract accounted for ₹4,000 and the other for ₹54,000. Their combined ₹58,000 across ten MWh gave ₹5,800 per MWh. An editor argued that the equal mean was simpler to explain and that changing it could suggest the source prices themselves were wrong. That concern helped identify what the correction needed to say. The original arithmetic remained valid for an equal-contract question. We had misdescribed its purpose, and I needed to own that distinction rather than replace one unexplained figure with another.

I authorised a review of the summary's interpretation and its publication treatment, working with the methodology and editorial owners. The correction retained the two rates, quantities, common period and stated energy-only scope. We explained both possible questions and identified which one the published label had promised to answer. Review and correction expenditure was ₹43 lakh. We did not equate the ₹800 difference between the two averages with a customer loss, an actual change in bills or a return from the review. No real tariff or purchasing recommendation was involved.

The episode changed what I required before approving a concise research figure. The intended population and unit now had to accompany the proposed calculation in the decision account. That did not make volume weighting the preferred method for every future study. An equal-contract average can be appropriate when the question genuinely concerns contracts with equal influence. My failure was allowing convenience to select the question without saying so. A useful correction restored that choice to qualified research discussion and made my own approval part of the published explanation.

In 2022 our current series contained revisions to seven of twenty-four monthly observations. Seventeen remained unchanged. The product team wanted a single clean download containing the latest values. That would make the current research easier to use, but a subscriber returning to an earlier publication could no longer establish what had been available when that publication was issued. The problem concerned the history of the research account, not a claim that every earlier observation was careless or that every revision improved a forecast.

I proposed retaining identifiable first-release and latest series with dated revision explanations. A product colleague objected that several versions could confuse readers and increase support work. That was a reasonable concern: a folder of unexplained extracts would preserve files without making their relationship understandable. We therefore needed a release structure that distinguished the recommended current account from the historical one, including which seven observations changed and the reason recorded for each change. The seventeen unchanged observations remained part of the same twenty-four-month comparison.

The programme carried ₹61 lakh expenditure for versioning and release work. Research owners assessed the reasons and implications of revisions; my team preserved their decisions in the release account. Readers could obtain the current series while also recovering the first release under its date. We avoided presenting today's corrected values as though an earlier reader had possessed them. We also avoided treating the retained history as a licence to continue distributing a known error without explanation. Both the correction and the sequence of publication needed to be intelligible.

The resulting product offered a more defensible account of what changed and when. It did not demonstrate forecast skill, quantify a trading opportunity or validate decisions made by subscribers. Those claims would require separate methods and evidence. Nor did retaining a version settle whether a particular revision reflected better coverage, a changed definition or an error; its explanation still mattered. The lesson for my role was practical: the current value and the value previously communicated are different research facts. A publisher should help readers distinguish them without requiring access to an employee's private archive.

In 2023 a research paper used responses from 120 industrial participants. Eighty-four were urban-industrial respondents and thirty-six belonged to the remaining defined panel categories. The draft introduction described the result in language that could be read as a national industrial estimate. The panel had not been recruited from a representative national frame. Complete responses and a clear spreadsheet did not resolve that limitation. I raised the issue before the wider claim became the basis of a subscriber-facing account.

The commissioning sponsor argued that the respondents were experienced and that a narrower description would reduce the paper's commercial appeal. Their experience did make the responses useful for questions about the participating group. It did not establish how the panel related to industrial users outside it. I asked the research team to identify the population the study could actually describe, the recruitment route and the coverage absent from that route. Applying convenient weights to the existing respondents would not, by itself, supply the missing design or prove that unobserved groups behaved similarly.

We narrowed the publication to an account of the defined panel and retained the composition alongside its findings. A ₹47 lakh programme examined recruitment design and the evidence a broader future study would require. This was research and design expenditure, not a claim that the original 120 had become representative. We did not invent an enlarged sample or report a revised national estimate. Methodology colleagues retained responsibility for design choices, while the commissioning and commercial owners had to explain the narrower product honestly to its intended readers.

The immediate consequence was less expansive language and further work before any broader claim could be considered. The panel remained valuable within its stated scope, including the questions its participants could illuminate. I did not want an audit of limitations to imply that all selected-panel research was useless. The leadership decision was to align the promise with the evidence we actually held. A reader should be able to see whose experience informs the account and whose experience remains outside it, instead of discovering that distinction only after attempting to apply the conclusion elsewhere.

In 2024 an external research package was being assembled from eighteen licensed data feeds. The delivery schedule assumed that technical access made each feed available for the package. Our review found twelve limited to internal analysis, four permitting defined redistribution and two whose proposed use still required interpretation. Those were different permission states. A functioning import process could not convert them into one unrestricted product, and a derived chart did not automatically settle the rights attaching to its underlying material.

The product sponsor argued that customers would receive analysis rather than raw files and that withholding material would weaken the launch. I understood the product concern, but the actual terms required qualified legal and licensing review. I would not offer my own general legal ruling about what every derived output allowed. The task was to describe the intended use precisely enough for the appropriate owners and licensors to assess it. We needed to distinguish internal research, redistribution and unresolved interpretation instead of treating a broad product label as permission.

The review cost ₹58 lakh. The external package used the four feeds within their verified scope. The twelve internal-analysis feeds retained their restrictions, and the two unresolved feeds were withheld from the proposed external use while interpretation continued. My team connected the source conditions to the package components and release decisions. Legal and licensing owners retained authority over interpretation; commercial colleagues retained responsibility for explaining the resulting offer. We did not describe the excluded material as unusable for every purpose, because the review concerned particular conditions and uses.

The package therefore had a narrower supported source base than the initial schedule assumed. I did not turn the decision into a quantified avoided fine or assert that a legal dispute had been prevented. Neither outcome was established. The useful result was a release account that matched the permissions actually reviewed and left unresolved questions open. Data leadership contributed by making those conditions travel with the product design. It could not replace the contracts or the people qualified to interpret them, even where a broader dataset would have made the launch more attractive.

In 2025 a research comparison moved to a revised industrial classification. Of sixty respondent firms, forty-four retained stable categories, ten involved categories that had split and six had ambiguous attribution. A simple conversion would have forced all sixty into the new labels and produced apparently complete comparisons. Yet a changed category could alter the reported distribution without any physical change in industrial activity. We needed to separate what the new classification described from what the firms had actually done.

The publication team wanted continuity in the familiar chart series. Leaving unresolved cases visible would complicate the presentation and reduce the set supporting some new-category comparisons. I regarded continuity as a valid reader need, but not a reason to invent attribution. We examined the supporting descriptions for the ten split-category cases and retained the six ambiguous ones for further work. The question was whether evidence supported a mapping under the new definitions, rather than which destination would make the historical chart look most stable.

The reconstruction programme cost ₹73 lakh. Seven of the ten split cases gained supported mappings; three remained unresolved. Together with the forty-four stable cases, that produced fifty-one supported mappings and nine unresolved records, including the original six ambiguous cases. We retained the old and new labels with their dates and mapping basis. Unresolved firms were not deleted from the research archive. They were excluded only from comparisons that depended on a supported new-category attribution, with the resulting scope made explicit to the reader.

The outcome was a comparison with a visible boundary, not a complete conversion of every respondent. Methodology owners still had to assess whether a particular longitudinal interpretation was appropriate. We reported no reduction in consumption, emissions or expenditure from changing labels, because classification itself established none of those physical or financial outcomes. The case reinforced a responsibility that sits between data processing and executive communication: when a series appears to move, the account should say whether the movement reflects observed activity, changed definitions or an unresolved combination. A tidy category field is useful only if its meaning survives the comparison it is asked to support.