Never silently change a number someone gets paid on

A driver's mileage tracking is crucial because it directly impacts their earnings. However, an algorithm might sometimes adjust the recorded distance, leading to potential underpayment. It's essential to maintain transparency in this process. The algorithm removes various types of distance readings, including those from mock location apps, physically implausible jumps, and sudden jumps over five kilometers.

The modified distance is more accurate than the raw sum, but the mistake lies in assuming accuracy is the sole objective. A driver who believes they covered 40km but sees 36km cannot differentiate between a meticulous algorithm and a covert expense reduction. From their perspective, both scenarios appear identical: a smaller number without explanation.

Accuracy, while important, is not the entire job. The algorithm's output should be visible to the driver, allowing them to understand the original distance and the proposed distance after adjustments. The algorithm's changes should be explicitly labeled in the user interface, showing the original distance next to the cleaned figure.

Both numbers should be present in the data model, so displaying them both is feasible. The user should be informed about the removed categories, such as abnormal distance, mock distance, and spike distance. The categories should be easily identifiable, unlike vague explanations. Users should have the option to override the algorithm's decisions.

This is possible because the driver was physically present, and they possess information that the algorithm does not. If the algorithm's threshold is incorrect, this toggle provides the quickest feedback mechanism available. The alternative would be for the driver to escalate the issue to their finance team, who would then escalate it to the company, three weeks later, without any supporting data.

An app-killed event, which the tracker recovers from automatically, should not be classified as an irregularity. If every notice is alarming, users may dismiss them all, rendering the disclosure ineffective. The data model must support this transparency. If the pipeline discards what it rejects, transparency becomes impossible. Naming the categories and allowing users to toggle removal requires the removed distance to still be present in the data model.

Transparency is a data-modeling decision made earlier, not a UI decision. The principle applies to various systems where the output becomes an input for someone's money, time, or reputation, such as fraud scores, ranking systems, time trackers, and usage-based billing. In all these cases, correctness is fundamental. Being right silently is insufficient; showing the working is how you earn trust.