Our pipeline reads sustainability reports and extracts environmental figures from them. Early versions trusted whatever a report said. The current version discards some numbers that appear, accurately quoted, in official company documents. This post explains why, because the reasons say something about how these reports are written.
The division-level renewable claim
Samsung's sustainability reporting headlines a renewable energy figure of 93.4% for its DX (consumer electronics) division. The figure is true. It is also not the company: Samsung's company-wide renewable share is roughly 33%, because the DX division excludes the semiconductor fabs that consume most of Samsung's electricity, and those fabs run largely on the Korean grid, where renewable supply is limited. Samsung's overseas sites run near 90% renewable; Korea, where the load is, does not.
A reader skimming the report takes away "Samsung: 93% renewable." So did our extractor. The pipeline now cross-checks every extracted renewable percentage against a researched company-wide figure, and a disagreement of more than 40 percentage points means the extraction grabbed a division-level or goal figure, which is then discarded in favor of the company-wide number. The pattern is not unique to Samsung; headline placement of the most flattering organizational slice is common across the reports we process.
The energy number that fails physics
The second guard is a capacity cross-check. Given a company's known data center capacity, expected consumption is capacity times 8,760 hours times a utilization factor times PUE. When an extracted energy figure deviates from that expectation by more than a factor of three, we discard it and estimate instead.
The figures this catches are rarely fabrications. They are industry-wide statistics quoted in a company's own report ("data centers consumed 415 TWh globally"), prior-year or target numbers, and unit errors, all of which an automated reader, or a hurried human one, will happily attribute to the company itself. Our own correction log shows what happens when a number like that slips through.
What this means for a B tier
A figure marked tier B on this site ("company-reported") is one that appeared in the company's own report and survived both checks. That is a deliberately higher bar than "the report says so." The full rules, constants, and their sources are on the methodology page.