Written by JD Davenport, who still sells for a living. Updated 4 November 2026.
A check a person runs is a checklist. A check that must return zero before a file saves is a gate.
Only the second one counts, because the first one fails on the day everybody is busy, which is the day the bad file ships.
That is not a comment about discipline. It is a statement about when failures happen. Checklists are skipped under load, and load is exactly the condition that produces the errors the checklist was written to catch. The tool is least available precisely when it is most needed.
It runs without being invoked. Nobody decides whether today is a day for checking. It is wired into the act of saving, shipping or publishing, and the work does not complete until it passes.
It returns zero or it stops. Not a warning. Not a yellow flag on a dashboard somebody reads on Fridays. A number that must be zero, and a refusal if it is not.
It can tell wrong from absent. This is the part most people skip, and it is the part that makes a gate trustworthy. A gate that reports 23 checks run, 16 passed, 7 failed, and all seven failing because the file does not exist yet, is a gate you can believe. A gate that quietly counts a missing file as a pass is worse than no gate, because it produces confidence rather than information.
A duplicate check across a document corpus, run by content hash rather than by filename. That one reduced 802 files to 209 genuinely unique ones. Filenames had been lying about duplication in both directions.
A pre-save pattern check on formatting that must return zero rows. The rule says how to write it, the search proves compliance before the file saves, and nobody has to remember either one.
Checks that walk every part of a file rather than the visible values only. Cell comments, validation prompts, headers and footers, defined names, document properties. Text leaks into all of those, and nobody looks at any of them.
A configuration-driven name check. The check it replaced tested three fixed strings, which meant it could never have caught a fourth name. A check with a hard-coded list is a check with a published blind spot.
A filename is a rollup.
So is a folder timestamp. So is a side panel, a summary line, a thumbnail, a heading, a byte count and a search hit.
The rule that a total is not evidence was written for money, and it applies to every one of those without a word of modification. A folder that says it holds twelve files is making a claim. Opening the folder is the evidence.
So: count files you opened, not filenames you saw. And publish the coverage where the reader can check it, because a coverage figure nobody can audit is itself a rollup.
Pick the one error your business has made more than twice. Not the worst one. The most repeated one.
Then ask what would have had to be impossible for it not to have happened, and build that, rather than writing a reminder about it.
If the answer involves somebody remembering something, the fix is not finished. Go back and find the mechanism.
Most of [the ways a job's real cost goes missing](/faq/which-jobs-made-money/) are gate-shaped problems wearing a training problem's clothes.
Adapted from The Audit, "What a machine can check for you". More in the books. Other notes on the notes page.