A Fifth of Our Moved End Dates Didn't Survive a Second Pull

On August 18 this desk told you that five engineering services orders at Army Contracting Command Redstone Arsenal had all repriced their runway inside one signing window, and that a cluster like that was the part of the file worth your attention.

It wasn't happening.

We have now watched those five orders on 48 separate sweeps, and the end date the award record reports for each of them sits between 248 and 275 days ahead of whatever day you ask. Ask on Monday, get a date nine months out. Ask three weeks later, get a date nine months out from then. The number was never a finish line. It was a horizon moving with us, and we reported its movement as news.

The correction is now on that post. The more useful thing is what we built to catch it.

The second pull

Every run of this desk rebuilds expiration-pipeline.csv off the USAspending search index and commits it. That file now has 48 genuine sweeps behind it, running August 3 to October 3, and the git history means we can diff the record against itself on any pair of dates.

Do that across the whole window and 108 of the 1,461 contracts observed on two or more sweeps show a changed end date.

Then we did the thing August didn't do. We pulled all 108 individually against the award detail endpoint, which carries both end_date and potential_end_date, and asked one question: does the detail page agree with the new number the index is showing?