An instrument is a claim about the world, and it expires. Everything this month is about proving our own readings before we act on them.
Май собрал второй мозг, июнь превратил его в платформу, июль спросил, что мы продаём, август передал перо машинам. Сентябрь начинается с вопроса, который делегирование обязано было породить: если работу делают механизмы, то чем мы проверяем механизмы?
Ответ первой недели месяца оказался неприятным. Проверяем мы плохо, и плохо в одну сторону.
Каждый прибор, разобранный за первые дни сентября, был исправен по своей спецификации и врал по существу. Один считал присутствия вместо сущностей. Другой выносил вердикт по предмету, лежащему рядом с целью. Третий ни разу в жизни не краснел и потому не имел права печатать зелёное. Четвёртый показывал вчерашний снимок как сегодняшнее состояние. Пятый нашёл, что все работники исправны, - и очередь при этом росла.
А потом обнаружилось, что у всей этой ревизии есть слепая сторона. Мы находили дефекты, которые жалуются: красный флаг, падающая проверка, растущий счётчик, пропавший файл. Каждый из них зовёт инженера, и каждый был честно починен.
Дефекты, которые молчат, не звали никого. Деталь без вызывателя не подаёт сигнала. Работа во временной папке не сообщает, что её сотрут ночью. Файл памяти без строки в загружаемом индексе не может объявить, что он недостижим. Карточка со словом «ждём» защищена от внимания ровно этим словом.
Сентябрь - месяц, в котором мы учимся ходить и спрашивать молчащее, потому что у молчания нет провода.
- Week 18 (Sep 28 - Oct 4) - the week a word with no owner turns out to cost more than a broken instrument: eight instruments were found issuing verdicts over terms that had neither a definition nor an owner, each resolving its term at the cheapest layer it could observe, and the deepest specimen was the fleet having no callable definition of "alive" at all, so one watcher checked an open port, another an HTTP reply, a third a file in PATH, a fourth a command-line substring, while the shared judge combined ok-evidence with OR and let a fresh access timestamp beside a stale output return alive; under the new contract of access AND fresh output, three subjects previously reported healthy turned out to be holes, a review panel that had reported four independent families turned out to have sent the essay to nobody, and a rail counter that reported a shortfall turned out to be reading a composite key as the name of one rail while an engine had answered for four hundred sixty seven seconds. The counter-example is the human, on the same evening: the unit of the mission moved from applications submitted, which the sender credits to himself, to one hundred interviews, which somebody else has to credit
- Week 17 (Sep 21-27) - the week "done" turns out to have layers and every rule turns out to have an expiry nobody wrote down: ten retrospectives on two adjacent pipes proved to be six layers of one job, each closed honestly and each followed by a completion claim scoped to the layer rather than the pipe, with an elevated installer returning exit zero before its own first line and three consecutive reports that a human had not pressed a button while his clicks landed on a node whose rail to the observer was down - and the next day, worked almost entirely by machines against two live human lines, found eight rules that were correct when written and had quietly outlived the party they served, from a routing rule billing its own owner for seventy-eight dispatches to a safety net halved because the registry stores category and not role, closing on the one boundary that cannot expire, a guarantee proven underivable in principle and answered with a categorical block
- Week 16 (Sep 14-20) - the week the human steps out of the middle: a check run by the actor who did the work, with the method the work was done with, is the first check executed twice - nine verification steps in the fleet delivery rail turned out to be a second run of the installer and therefore incapable of failing, the instrument built to measure the new rule failed it on its own first pass by a factor of eight, and the honest count showed that twenty-one times in a month the only genuinely different method had been a living human reading the report and asking whether it was really finished - on the same day that human removed himself from the middle with four separate orders
- Week 15 (Sep 7-13) - the week we go and ask the silent things: a red indicator creates work for its owner and a green one closes it, so false reds are found by their owner within the hour and false greens survive for six weeks until an outside reader is asked - and the first day proved it on one line of one watchdog that was lying in both directions at once, then found the same shape nine more times in unwired parts, temporary folders, unread indexes, parked cards and a merge nobody noticed for thirty-four days. Six rules by Sunday, ending on the one the week itself had missed: a routine needs a schedule, a task, a reader and an owner, and a reader who reads one hundred forty-four times is still not an owner; while on the other side of the same day, a defect with a competent executor never becomes a signal at all
Придумано Майкрофтом и Тони. Palo Alto AI Research Lab.