Seven broken sentences reached print. 4 passes missed them

What happened

A retrofit, a structural edit that spliced new material onto existing paragraphs, left seven sentences cut in half mid-thought. The broken halves sat in the manuscript through four separate scoring and reading passes, and reached print.

What the numbers showed

Four scoring passes, each one built to check something specific: structure, receipts, voice, promise delivery. None of them were built to catch a sentence tail spliced onto the wrong paragraph. Every pass reported a result on the thing it was measuring, and every one of those results was accurate. The defect simply wasn't in any of their four lanes.

Why it failed

A reading pass built to judge whether an argument holds together will read past a broken sentence if the argument around it still makes sense, because that's what it's built to notice. Four passes checking four different things about the same manuscript is not the same as one pass checking punctuation, and nobody had built that fifth pass.

What we changed

A mechanical gate now runs automatically on every build, with no manual override, and this exact defect class is a hard failure inside it: no paragraph allowed to end without terminal punctuation, among other fixed, countable rules. The gate doesn't read for argument quality at all. It reads for the narrow, mechanical things the four scoring passes were never built to catch.

How to spot it in your own account

If your review process is a set of scoring passes judging quality, voice, or structure, check what none of them are built to catch. A gate that checks for missing terminal punctuation, broken sentence splices, or other purely mechanical defects costs almost nothing to build and catches a category of error that a quality-focused reading pass will walk straight past, because walking past it is exactly what a quality-focused pass is designed to do.