An exit code found 11 gloss violations in a three-line draft, 68 days after the style guide went unverified

Sixty-eight days produced no recorded violation because, by the team's own account on dev.to, nothing in the repository was looking for one ^[3]. The same instruction file already had a rule that could not rot: a byte budget enforced by a test, where the body growing past the declared number makes `pnpm test` fail, and the only way through is to move something out or raise the budget in the same commit with a written reason ^[7]. Byte counts are observable from outside the writer; whether a sentence is clear to a newcomer is not, and the post says you cannot diff a paragraph against comprehension ^[8].

There is a second reason the style rule sat idle. The author writes that a language model reads a rule like that and immediately believes it is already following it, because the model is the one entity in the system that does not need the gloss: it knows what `infeasible-at-cap` means, and it wrote the code that emits it ^[5]. The owner's second ask, 68 days after the first, is what the author treats as the evidence, on the grounds that anybody can miss a rule once ^[6].

The check that shipped is narrower than the rule it replaced. It stops at whether every internal term, on first appearance, sits next to a plain-language gloss, and the author says plainly that this does not prove the paragraph is clear ^[9]. What it catches is the failure the owner complained about, which was a run of `verdict`, `FOLLOW_DAILY_CAP` and `stock-runway` with nothing around them ^[10].

The checker accepts exactly three gloss shapes: term (gloss), gloss (term), and term = gloss ^[14]. A parenthetical counts only when it holds two or more consecutive characters of prose in the reader's language, sitting adjacent to the term ^[15]. The first implementation of the third form accepted an equals sign followed by any text containing prose within the next 24 characters. That passed `verdict=infeasible-at-cap was raised`, because prose appeared later in the sentence, and it recorded `verdict` as glossed. The gloss now has to begin immediately after the equals sign ^[16].

The checker holds a set of seen terms, so each one is reported once ^[17]. So the eleven violations in the three-line draft are eleven distinct terms ^[19], about 3.7 a line ^[18]. For that rate to say anything about your own drafts, your prose has to be as dense in internal vocabulary as a shop report whose nouns are command names and ledger fields.

The term list is derived from artefacts that already have to be correct, on the reasoning that a hand-maintained list of forbidden words would have rotted faster than the rule it replaced ^[11]. When the checker finds an unglossed term the glossary already defines, it returns the first 60 characters of that definition as the suggested fix ^[13]. A definition that has drifted therefore ships as the sentence the reviewer is told to write. The post does not report how many of the eleven flagged terms the reviewer accepted ^[21]. One row of the dictionary stays hand-maintained, because `behind` is an ordinary English word everywhere except in their ledgers ^[12].