The night shift knows things nobody wrote down

A plant wanted to extend unattended running from four hours to a full night. The machines were capable, the material handling was automated and the sensing was in place. Two attempts had been made and both were stopped before dawn.

Both stops came down to the same thing: something happened that was not in any procedure, and the system did the only thing it could, which was to halt and wait. On a staffed shift those situations are resolved in seconds by somebody who has seen them before.

What that knowledge actually is

Which machine drifts, and when. The third one runs hot after a long changeover and the first two parts should be checked. This is in nobody’s manual and everybody on that line knows it.

Which material behaves differently. A particular supplier lot runs slightly outside nominal in a way that is fine, and another one is fine on paper and causes trouble.

What a deviation means in context. The same reading is normal after a restart and a problem in hour six. Interpretation depends on state that the instrument does not have.

When to stop and when to continue. Most of the value. Experienced operators stop things that look acceptable and continue through things that look alarming, and they are usually right.

What has changed recently. A new tool, a different pallet supplier, a maintenance job last Tuesday. Context that never enters any system and explains most of what looks like random variation.

Why you cannot simply ask for it

Ask an experienced operator what they know and you get a modest answer. Not because they are withholding — because the knowledge is recognition rather than a list. They know it when they see it, and they cannot enumerate it on request.

The technique that works is to invert the question. Build a draft rule and get it wrong in front of them. People who cannot list what they know will correct a wrong version immediately and precisely, and the correction is the specification.

Real cases work the same way: take twenty deviations from the last quarter and ask what they did and why. The reasoning comes out attached to the case rather than in the abstract.

The taxonomy is never complete, and that is the design

There will always be a situation nobody anticipated. The important decision is what the system does then, and there is only one safe answer: stop and call somebody.

A system that guesses at an unrecognised exception at three in the morning is worse than one that halts, because a halt costs a few hours and a confident wrong decision costs a shift of scrap or a customer. Each of those calls becomes a new entry in the taxonomy, which is how the coverage grows — slowly, from evidence, rather than from an attempt to enumerate everything up front.

Where to start

Not with the robotics. With one line, twenty real deviations from the past quarter, and the people who handled them. That work produces a context model and an honest assessment of which parts of the operation are genuinely ready to run unattended — usually a narrower set than the ambition, and always a more useful answer than a plan built on the assumption that the machines are the constraint.

That is what dark factory context development consists of, and most of it is spent with people who work nights.