High Vibration Is Not a Severity

I spent an evening teaching a language model to read operator defect notes. It kept doing one thing wrong, over and over, and it took me longer than it should have to see that the model wasn't the problem.

Give it a line like "high vibration on the discharge pump" and it would classify the fault correctly — abnormal vibration, no argument — and then mark the severity Critical.

Every time.

01The machine was reading English correctly

That is the part worth sitting with. It wasn't hallucinating. It wasn't confused. In ordinary English, "high" is an intensifier — high risk, high cost, high fever. A high anything is worse than a low anything. The model applied the most reasonable reading available to it and produced a wrong answer with total confidence.

Because on a plant floor, "high vibration" isn't a judgement. It's a name. It is what the condition is called, the same way "low flow" or "high differential" are called what they're called. The word "high" is doing taxonomy, not triage.

A machine that reads your logs as English will read half your fault names as emergencies.

And the failure mode is nasty, because it is silent and it is one-directional. Nothing gets missed — things get promoted. Your critical list fills with Tuesday-afternoon observations. Then the one entry that genuinely was critical sits in a list of ninety other criticals, and the list stops meaning anything. You haven't lost a record. You've lost the ability to sort.

02Where severity actually lives

Here's the thing I knew from shift work and hadn't ever had to say out loud: on the floor, severity is never carried by the description of the fault. It's carried by a separate signal, and it's usually one short phrase riding alongside.

• "Tripped." "It's down." "Unsafe to run." — that's critical, and none of those words describe the fault.
• "Getting worse." "Needs attention this shift." — high, and note it's about a trend, not a state.
• "Keep an eye on it." "Monitor next round." — medium. This is the most common thing an operator actually says.
• "Minor." "FYI." "When you get a chance." — low, and worth capturing precisely because it usually isn't.

Read that list again and notice what's absent: not one of those cues says anything about what's wrong with the machine. Urgency and diagnosis are two separate fields, and a shift supervisor keeps them separate without ever being taught to. The system was the thing that collapsed them.

03The fix was one sentence, not a bigger model

What fixed it wasn't more parameters, a fine-tune, or a longer prompt. It was one instruction, written from the floor:

That last clause is the one that matters most, and it was the hardest to accept. The obvious instinct is to have the system guess — pick a default, put something in the field, keep the record tidy. That instinct is wrong. A guessed severity is indistinguishable from a judged one the moment it's stored, and a field full of confident guesses is worse than a field with honest gaps. An empty severity says nobody decided yet. A guessed one lies quietly, forever.

04Why this is the whole argument about AI on a plant floor

A frontier model has read essentially every maintenance manual ever digitised. It knows more failure theory than I will in my lifetime. And it has never once heard a supervisor at three in the morning say "she's been screaming all night but she's holding."

Manuals are written in engineering English. Shifts are spoken in plant English. They share a vocabulary and disagree about what the words mean. That gap does not close with scale, because the missing information was never written down anywhere to be trained on — it lives in the heads of people who use it forty times a shift and have never had cause to explain it.

Which is the useful conclusion, and it cuts against most of what gets said about this: the constraint on AI in an operating plant is very rarely the model. It's that nobody has yet told the model what the floor already knows. One sentence about the word "high" was worth more than any upgrade I could have bought.

So the honest question isn't whether AI can read your shift log. It's whether anyone has checked what it thinks your words mean.

Pull ten defect entries from your own system. For each one, ask where its severity came from — a real judgement someone made, or a word like "high" that got promoted on the way in. If it's the second, that field isn't data. It's grammar.