Chief Data Officer / India
The tidy answer was not always the supported one
Decisions & consequences
01Undoing a work-based merge of distinct recordings
In 2022 I approved a trial rule that consolidated supplied recording records when they shared the same work reference. The staging set contained 100 distinct recordings associated with 40 works. The output contained 40 rows, and the project summary treated the reduction as successful cleaning. A catalogue colleague demonstrated two different performances that our output had merged. The work relationship was accurate, but it was the wrong basis for asserting recording identity.
The delivery sponsor wanted to preserve the reduced output and place performance differences in a note. That would have kept the planned display and avoided a model revision. I rejected the proposal because the recording distinction belonged in the structure used by the service, not in optional prose that downstream handling might omit. My approval had helped create the problem. I stopped customer delivery of the staging output and funded ₹72 lakh to restore distinct recording entities and their work relationships, with catalogue specialists examining examples that the original rule had mishandled.
The repaired set retained 100 recording records linked to 40 works. It did not create 100 new works or claim that every supplied identifier was a certificate of ownership. The erroneous output had remained in staging, so we did not describe the repair as recovery from a confirmed customer or royalty loss. The cost and delivery delay were real consequences of correcting a rule before it travelled into service. We retained the original trial evidence so that the review could distinguish our changed model from an unexplained increase in record count.
I changed the acceptance discussion from how many rows had disappeared to which identities and relationships remained faithful to the supplied evidence. A smaller catalogue can be a valid result, but row reduction alone cannot establish that the removed distinctions were redundant. I would now require challenging examples before a model decision receives senior approval. The lesson was particularly uncomfortable because the conceptual difference was familiar to our specialists; the failure was allowing a convenient project measure to override the question they were qualified to answer.
Keeping overlapping permission claims outside a supported service set
A 2023 review examined 120 recording-territory-period entries proposed for a client service. Ninety-two had documented scope supported by the responsible sources. Twenty-eight contained overlapping claims that could not both be treated as settled on the available evidence. The data was sufficiently complete to fill the fields, but that completeness concealed disagreement about what the service could represent as permitted.
The account sponsor preferred to list both claims and let users decide which to rely on. That preserved a broader apparent offering and avoided holding entries while clarification continued. Devika and I considered that unsuitable for a service whose supported set would be interpreted as having a resolved scope. We funded ₹58 lakh for a bounded evidence review with authorised client representatives and the relevant advisers. The data team organised the conflicting evidence and its territory and period; it did not select a legal owner or rewrite the underlying agreements.
Nineteen of the twenty-eight questioned entries received clarification from the responsible sources. Nine remained outside the supported set. Together with the original ninety-two, that left 111 entries with a documented scope for the service under review. The nine were not deleted from the evidence record or described as permanently unusable. They remained unresolved for this proposed handling until the appropriate owners could provide a supported instruction. No finding about all rights in the underlying recordings was asserted.
The commercial cost was a narrower service and continued work on exceptions, alongside the retained ₹58 lakh review expense. We did not claim recovered royalties, prevented litigation or a financial benefit measured from the hold. I would now ask how the product distinguishes an entry that exists from one it is prepared to represent as supported. Directors should be able to understand that distinction without being asked to resolve the rights dispute themselves. A data function earns credibility by making the boundary visible, including where a customer wants a more complete-looking answer than the evidence permits.
Separating contributor roles before calling a credit complete
In 2024 we examined 240 supplied credit rows for a catalogue service. One generic creator field contained names whose underlying roles differed. A person could be a performer, composer or producer, and the supplied name alone did not tell the receiving system which relationship was intended. One hundred and seventy-six rows were confirmed as stated; sixty-four required clarification. The problem was semantic rather than a spelling or capitalisation defect.
The operations sponsor proposed filling the ambiguous rows from the most prominent displayed artist name so that the release could proceed with a complete creator column. The attraction was a faster and visually consistent catalogue. I declined that substitution. Prominence was not evidence of the role the field was supposed to describe. We allocated ₹39 lakh to work with responsible catalogue sources, retaining the original description and requesting the specific role information needed for the receiving service.
Forty-six questioned rows were clarified. Eighteen remained held, producing 222 confirmed role descriptions from the original 240. We did not turn the held entries into blank credits without explanation or treat the unresolved person as having made no contribution. The issue was our ability to represent the role accurately in this service. Nor did confirmation establish a royalty allocation: a credit role and a payment entitlement are different questions with their own evidence and responsible parties.
The decision required a less uniform release account and more work by specialists whose contribution could be hidden in a simple completion percentage. I explained to Priya why the eighteen unresolved rows needed to remain visible rather than be counted as a failure the team should conceal. The outcome was a more defensible description of the supplied credits, not a claim of universal attribution accuracy. I would now agree the role vocabulary and evidence requirements before promising that a generic name column can serve a more specific purpose. An attractive display does not compensate for assigning a person a contribution the source has not established.
Replacing a single-language default with four meaningful states
A 2025 catalogue exercise required a language description for 200 items. The receiving screen expected one language value for each item, and the proposed completion rule would have filled uncertain records from the language associated with the supplying catalogue. Listening and source review showed that this convenience did not describe the material reliably. Some items used several languages, some were instrumental and some lacked enough evidence for a supported assignment.
The product sponsor preferred a single value to preserve an existing filter and avoid explaining exceptions. I accepted that a simple filter could help users, but not that its convenience justified manufacturing certainty. We funded ₹44 lakh for catalogue and language-practitioner review and a revised internal representation. The work retained one supported language where appropriate, allowed several where the item called for them, and kept instrumental and unknown as distinct states. It was a service-specific description, not a claim to classify every cultural or linguistic relationship in music.
The reviewed set contained 112 items with one supported language, thirty-eight with multiple languages, thirty instrumental items and twenty unknowns. Those four groups accounted for all 200 items. Unknown did not mean instrumental, and multiple languages did not mean a defective record awaiting reduction to one. Product colleagues had to reconsider how the filter and display would communicate those distinctions. We did not assert a measured improvement in discovery, listening or revenue from the change.
The trade-off was a less convenient implementation and a visible limit to the evidence. I regarded that as preferable to a completed column that users would reasonably read as an informed description. The team retained the source basis for assignments and a way to revisit unknowns when further evidence arrived. I would now involve catalogue and language expertise before finalising the field that will constrain their answers. The model should permit a truthful description of the material, rather than make specialists choose an inaccurate answer because it is the only answer the screen accepts.