The check I asked for would have passed the work I rejected

On Tuesday morning I rejected a LinkedIn document my own pipeline had built and told it to try again. Too wordy, I said, and asked for a twelfth-grade reading level. The deck had already cleared four gates on its way to me.

This afternoon I ran the gate I built that day against the archived version of the deck I threw out. Flesch-Kincaid grade 3.4, against a ceiling of 12. Then I ran it against the rebuild I approved and shipped. Also 3.4.

The number I asked for would have passed both of them, with more than eight grade levels to spare, without registering that anything had changed.

What was actually wrong with it

The gate returns a fifteen-item fix list on the rejected deck. Not one of the fifteen is about reading level.

Seven body slides ran over the twelve-word cap, one at twenty-five words and one at thirty. Slide two was the two-word fragment "What did", which is not a second hook but the middle of slide one's sentence. Slides three and five were fragments too. The payoff number sat on slide seven of nine while the last card carried no number at all, which inverts the only mechanic the format has. The last card read "Close it at the source.", a bare imperative where the house rule asks the sub line to invite. The post body under the deck ran 252 words and came down to 205.

Set that list beside what I actually said. Seven of the fifteen findings are word counts, so I had the fault in view; what I got wrong was the instrument. Flesch-Kincaid counts syllables per word and words per sentence. My cards were short sentences of short words. There were simply too many of them on each card, in an order that gave the ending away on slide seven.

Reading level was the measurable-sounding phrase I had to hand for "this is hard to take in". It is not the same quantity, and if I had shipped it as the gate I would have built a check that passes the exact artefact that caused me to ask for a check.

A note is not a gate

Eight days earlier I had written this down already. A note I wrote on 24 August records that the humanizer, the AI-writer detector and the metadata strip measure none of jargon density, reading level or slide structure. The fix agreed that day was to add a plain-English pass by eye.

That fix lasted eight days and then the same class of unit came through the chain with thirty words on a card.

An eye pass is a gate staffed by whoever happens to be reading, in whatever state they are in, at whatever hour the run fires. It produces no output. When it lets something through you cannot tell afterwards whether the rule was wrong, the reader was tired, or nobody performed the pass at all, because all three leave the same trace, which is none. The second miss had the same cause as the first, and I had already written the cause down.

The gate was wrong within the hour

Pointing the new script at a second deck the same afternoon broke it three times.

It treated any digit on a middle card as the withheld payoff, but that deck legitimately needs "three mandates" and a date in August on its middle cards, so the rule now takes an explicit payoff key and enforces that one number rather than every numeral. Its fragment test was a hand-written list of verbs, which fired on the words "named" and "ties" used as ordinary past tense, and became a shape test instead. And "so" was in its list of connectives that mark a card as continuing the previous one, which meant it flagged the header of the deck I had just approved and shipped, "So I stopped writing workarounds".

That third one is the useful failure. A check whose first act is to condemn something I had already judged good told me the rule was too crude on day one, for the price of an afternoon, off a deck I could verify by hand.

The number moved and I cannot say exactly why

The record written on Tuesday says the rejected deck returned a thirteen-item fix list, and breaks it down as four objections plus nine breaches of the word budget. Running it now, it returns fifteen, of which seven are word budget and eight are structure. Three patches went into the gate between the two runs. Both halves of that list moved, in opposite directions, and I have not reconciled them item by item.

I kept the four decks, two rejected and two shipped, and called them a regression suite. What I actually kept was four inputs. I recorded no expected output for any of them, so the suite can tell me a rejected deck still fails and a shipped deck still passes, and nothing about whether either one fails the way it failed on Tuesday. The list changed shape between one afternoon and the next, and the only reason I know is that somebody had written the number thirteen into a sentence.