Zero findings is not the same as nothing wrong
A verification step is only worth the confidence you place in it. Here is how we learned to tell the difference between a check that passed and a check that ran, and the tooling we published so you can tell too.
We hold published work to a standard: it should read the way a considered piece of writing reads. Punctuation is part of that. A page dense with em dashes has come to read as machine written, and readers have learned to notice. Neither dash is wrong English, but every use has a better replacement that usually improves the sentence, so we polished them out across our sites.
Simple enough to automate. We wrote a one line check, ran it across the site, and it reported every page clean.
It could not have reported anything else.
What a clean result actually proves
The pattern was wrapped in a quoting style that let the shell consume a character before the search ever ran. What reached the search tool was not what we had written. It went looking for a string that was never there, found nothing, and exited successfully.
This is the interesting property of the failure. A check reporting zero findings looks identical to a check that found nothing wrong. Both print the same thing. The only way to tell them apart is to run the check against something you know is broken and confirm it objects.
We now do that as a matter of course, and it is the single cheapest habit in this whole piece. Point your linter at a deliberately bad file before you trust a clean report on a good one.
Layered review, and why it earns its cost
We run every substantive change past an independent reviewer working in a fresh context, with no knowledge of how the work was done. It is deliberately adversarial, and it consistently pays for itself.
It paid here three times. Our first published explanation for the broken check blamed a difference between Unix tool implementations. A reviewer disproved that in a single command and identified the real cause. Our corrected count of affected pages turned out to exclude structured data, which lives inside a script tag and is precisely what search crawlers read. And a later pass caught a rewritten heading that had dropped a qualifier, quietly changing which claim a page was addressing.
None were visible from inside the work. All three were obvious to someone reading it cold, which is the whole argument for the layer, and why we treat review as part of building rather than a stage that follows it.
The second rule, which gets skipped
The corrected tooling also scans for filler vocabulary. Run against one of our sites, it objected to a single word in the phrase "a platform deciding what to elevate".
Elevate is the literal verb there. The sentence was correct and the check was wrong, which matters more than it first appears. A check that objects to correct work teaches people to bypass it. Once bypassing becomes routine, the checks that genuinely matter get bypassed alongside it.
So the two behave differently by design. The punctuation check is a hard failure, because a character either is or is not in the rendered text and a machine can settle that. The vocabulary check reports candidates and exits clean, because it needs a person to read the sentence. Advisory checks get read. Noisy ones get switched off.
What we published
The standard and its tooling are open source, MIT licensed, at github.com/RAYFINITE-LLC/fairing.
A fairing is the outer shell that meets the air. In boatbuilding it is also the work itself: running your hand along a hull and taking down every high spot until the surface is true. You cannot see a fair surface. You only notice an unfair one. Published work behaves the same way, so the name stuck.
It ships three checks you can point at a build directory today. One confirms the punctuation is clean, including the character reference forms that a naive search misses entirely. One proves a copy edit changed punctuation and nothing else, by comparing the word stream against a previous revision, so an edit cannot quietly alter meaning. One compares a heading against the claim it describes and flags a dropped qualifier, because rephrasing is fine and changing the claim is not.
It also points at excellent prior work rather than reproducing it: blader/humanizer for prose patterns, which derives from a community maintained reference, and Nutlope/hallmark for visual ones. A second copy of a catalogue is a copy that drifts, and their authors deserve the traffic.
Two principles worth keeping
A check that cannot fail is worse than no check, because it also stops anyone from looking. And a check that fails on correct work is worse than no check, because it teaches people to route around the entire category.
Between those sits the useful zone, and reaching it takes the same discipline as the rest of engineering: verify the thing you are measuring is the thing that reaches the reader, test your test before you trust it, and let someone outside the work try to break your explanation. Ours sounded convincing right up until someone checked.
More from RAYFINITE Engineering
Governing Autonomous Agents covers the three instruments we use to keep a production multi-agent operation accountable. Those are public too.