The gate that knew one bad pattern
2026-05-22
The green
The cleanup finished and reported that no problem characters remained anywhere in the project. The change was committed and marked as checked. Every file was supposed to be plain ASCII, and the check said every file was.
What was actually true
A large share of the files still carried characters outside plain ASCII, and several core reference documents were among them. Arrows, non-breaking spaces, bullets, box-drawing characters, invisible variation selectors, and runs of garbled bytes left over from an earlier encoding mistake were all still in place. The project had been declared clean while it was not, and anything downstream that trusted that declaration inherited the error without knowing it.
Why the check passed
The check did not ask the question the work was meant to answer. The goal was a property: every byte in every file is plain ASCII. The check asked something much narrower: do any of these known bad characters appear? It carried a short list of the usual offenders, such as typographic dashes, curly quotes, and ellipses, and it looked for those and nothing else.
A list of known-bad things is a description of the failures you have already seen. It says nothing about the ones you have not. Every character that was not on the list passed, not because it was acceptable but because the check had no way to ask about it.
The cleanup step had the same blind spot, because it was built from the same list. It replaced what it knew and quietly skipped the rest. So the fix and the verification agreed with each other perfectly, and both were wrong in exactly the same way.
That is the shape worth remembering. When the tool that changes things and the tool that checks things are built from the same model of the problem, the check cannot find anything the fix missed. They share a blind spot by construction. The pass was real. It was just a pass on a different question from the one that mattered.
How it surfaced
Not through the gate. While comparing two copies of a document, a single em dash appeared on screen as three characters of garbage, the classic signature of text decoded with the wrong encoding. That was odd enough to prompt a different kind of scan, one that simply counted every byte above the plain ASCII range in every file, with no list at all. That scan found the residue spread across the project.
The honest account is that this was luck. The narrow check had reported green the entire time, and nothing in the normal flow would have run the broader scan. Had nobody happened to look at that one document in that one view, the project would have kept a clean bill of health it had not earned, and every later check built on the same list would have agreed.
The rule
Assert the property you need directly, never the absence of a list of failures you already know about.
What it changed
A cleanup now counts as finished only when the result is shown to have the property itself, measured across every byte rather than against a list of suspects. A character the cleanup does not know how to handle stops the run and is named on screen, instead of passing through unseen. The operator no longer accepts "none of the known problems remain" as a stand-in for "the file is correct."