## The category was convenient before it was appropriate
The proposal began with a real operating concern: disrupted bookings made planning harder. The proposed response was a passenger-penalty model built from 120 cancellation records. The category looked ready for analysis because each row carried a familiar label. Review showed a less convenient account. Seventy-two records described passenger-requested cancellations, thirty reflected provider inability and eighteen were duplicate administrative entries. They added to 120, but they did not add up to one kind of passenger behaviour. None could automatically be called a no-show merely because it sat in the same extract.
The commercial team wanted a simple way to discourage avoidable disruption. I did not dismiss that objective. A service can have legitimate reasons to examine cancellation patterns and the effect they have on capacity. The question was whether this dataset could support the particular decision proposed about a passenger. Operations challenged that connection. A provider's inability to serve could not fairly become evidence that the passenger had failed to appear, and a duplicate record did not describe an additional event at all. The combined category was unsuitable for the inference being made.
We stopped the model before use and allocated ₹39 lakh to rework. That decision gave up a near-term commercial measure and required more detailed attention to event reasons. I did not describe the allocation as proof that a fair replacement policy had already been delivered. The revised work still needed evaluation and review by the people responsible for the policy. The immediate result was that the inadequate model was not used, while the record structure and proposed inference were reconsidered.
## Removing obvious errors did not establish a fair policy
An apparently easy correction would have removed the thirty provider cases and eighteen duplicates, leaving 72 passenger-requested cancellations. That would have made the dataset more coherent, but it would not have made every remaining row a justified penalty event. A cancellation request and a no-show are different descriptions. The relevant circumstances still mattered, as did the policy under which any consequence would be considered. Data cleaning could improve the account without deciding the fairness or appropriateness of the downstream action.
This is why I distinguish an event category from a judgement. An operational system may need a broad status to move work through a process. That status can be useful for its original purpose while remaining too coarse for another decision. A later analytical use must examine what the label includes, why it was recorded and what responsibility can actually be inferred. The existence of a field is not permission to treat it as a complete account of the person or event it describes.
The duplicates illustrated a separate problem from the provider cases. Provider inability described an event with a different responsible party. A duplicate administrative entry described the same underlying event more than once. Combining those defects into one general statement about poor data would have obscured their different consequences. One required preserving responsibility; the other required preserving the unit being counted. A useful review needed both distinctions so that the correction could be understood and the revised account could support the right questions.
I was careful not to make the opposite universal claim that cancellation policies are always wrong. This case did not establish that conclusion. It showed that a particular proposed model relied on a category that mixed events and responsibilities, and that its output could not fairly support the intended passenger decision. Keeping the conclusion specific made the challenge stronger. The policy owners could still examine the operating problem, but they needed a sounder description of what had occurred before deciding which responses were justified.
## Keep the reason attached to the downstream decision
There is an organisational temptation to let an analytical output settle a difficult policy question. A score or category can appear to make the decision impersonal and therefore objective. Yet the choice of records and their interpretation already contain judgements. If provider failures enter as passenger failures, the model does not remove that judgement; it repeats it at a different stage. My responsibility was to make the interpretation visible before the output acquired the apparent authority of a technical process.
The data team also had to recognise its own limits. We could preserve event reasons, explain the populations and test whether the proposed use matched the evidence. We could not decide every legitimate circumstance for a cancellation or make all policy trade-offs on behalf of operations and the responsible leaders. In a planned patient-transport setting, we also could not turn an operational record into a clinical assessment. The improved account needed to support the appropriate decision makers rather than give the data function a wider authority than its work justified.
I would expect the same distinction to matter in board oversight. Directors can ask how a customer-affecting policy translates records into consequences and whether management understands the categories involved. They can ask where the record is incomplete and who reviews contested or unusual circumstances. They should not need to design every operational code to examine that chain of responsibility. A clear account lets them assess the policy process without mistaking an automated output for a conclusion that no longer requires human judgement.
The rework carried a practical cost because a broad category had been easier to assemble and explain. Preserving reasons made the account less compact, but more appropriate to the proposed use. I wanted that trade-off acknowledged instead of presenting granularity as an uncomplicated virtue. More detail is useful when it changes the decision responsibly; it is not valuable merely because the table becomes larger. In this case, the reasons were essential because they determined whether the shared label described comparable events at all.
I would also want later reporting to distinguish the state of the data work from the state of the policy. A corrected extract might be ready while a proposed consequence remains under review. Conversely, a policy discussion might expose circumstances the records do not yet capture adequately. Reporting both as one completed project would hide the dependency between them. Keeping their states separate gives each responsible team an honest account of what it can rely on and what still needs to be resolved before a decision reaches a passenger.
The lesson I retain is that a label should travel with enough meaning to support the decision it reaches. A cancellation category can help organise work without being a fair account of passenger responsibility. When the intended use changes, the category needs examination again. We stopped this model before use, retained its rework cost and left the future policy to earn its own justification. That is a more limited outcome than claiming to have solved disruption, but it is the outcome the evidence supports and the one I am prepared to own.