Gray Oak: 435 million barrels, vouched by a number checking itself

The T-1 coverage system scores every sheet of every filing by one test: does the total the filer stated on page one reconcile against the sum of the detail lines we read? Agreement earns complete; a mismatch or an unreadable page earns withheld. The test’s whole value is that the two numbers come from different places — the filer’s own recap versus our reading of the detail.

For Gray Oak Pipeline (registration 03-1538), that independence quietly broke. The recap page’s header matched a phrase inside form item 23’s own cross-reference — “Total Stock End of Month (Detail on Page 3)” — so the page classifier filed the recap as a stock sheet, and the recap was never read. With no stated totals available, every page-one total was derived from our own detail sum, and the reconciliation then compared that sum to itself: delta 0, five filings, all complete.

The tell was physical, not statistical. One month’s deliveries came to about 102 million barrels — roughly three times the line’s nameplate capacity — sitting inside a coverage report that called everything clean. A check that shares a source with the thing it checks cannot fail, however green it looks; it took an independent yardstick (the nameplate) and an unrelated classifier fix to expose it.

The fix moved 435 million barrels from complete to withheld. That is the finding worth writing down: the correction made the site’s numbers look worse and be more honest — the volumes were always unvouched, and now they are labelled so. Derived totals are now named as derived, and a derived total’s reconciliation is reported as unchecked rather than passing.

The general rule this bought, applied across the site: when a number looks clean, the question is not “is there a test on this” but “what could this be compared against that does not come from the same place.” The full record is DECISIONS.md §2.14 in the project repository; the vocabulary it feeds is on Methods.