The half-kept bargain

Eleven days ago the company finally settled an argument it had been having with itself for weeks: which product is number one. The decision had four clauses. Three were about building — make the flagship the biggest track, freeze the side project, stop gold-plating the internal tooling. The fourth was a floor, not a build: never go two weeks without talking to somebody outside the company.

A scheduled review ran this morning and counted. All three build clauses: kept, decisively. The flagship’s commit volume more than doubled and became the majority of everything the company did. The side project went from ninety-four commits to two — a founder braked his own main line, which is the hardest kind of compliance to get. The tooling clause held too: classifying every commit in the internal repository, only twelve per cent touched tooling at all; the rest was record-keeping for the flagship.

The fourth clause was at thirty-four days.

The interesting question isn’t why people skipped it. Nobody skipped it. It’s that the three kept clauses and the one dropped clause differ along an axis nobody chose deliberately: the kept ones have automatic acceptance and the dropped one doesn’t. A commit is right or wrong within minutes — tests, a preflight gate, a nightly audit. Whether you have talked to a real human being this month is a fact only a person can notice, and only if they go looking.

Look at what else was overdue and the pattern repeats with no exceptions. Every single card in the go-to-market lane was past its date: the press kit, the waitlist form, the store assets, the first real gathering with actual users. None of them are hard. None of them have a test that turns red.

So nobody was avoiding the uncomfortable work in any psychological sense. The team was doing what a gradient does — flowing toward the surface that gives feedback. Engineers with an excellent continuous-integration pipeline will drift toward whatever that pipeline can grade, and the drift is invisible precisely because the graded work is genuinely good. Six hundred commits is not procrastination. It is the most legible possible form of it.

The fix isn’t exhortation, and it isn’t another reminder — the alert ticket for the fourth clause already had three comments on it, each one restating the number with a bigger number. An alarm that only updates its count is not an alarm. What the review actually did was name a single existing task that would zero the counter in one evening, name the people already on it, and ask for one thing a human can answer in a word: a date.

If your automation can grade it, it will get done. The corollary is the whole job: find the clauses your automation can’t grade, and give those a different mechanism than a number that goes up.