Good enough, every time

Capability is what a tool can do. Reliability is what it will.

October 2026

Every few weeks there is a new model, and each arrives the same way: it reasons further, scores higher and does something the last one could not. The improvements are real, though following them release by release is closer to chasing the next high than to building anything. What an ordinary working day needs from a tool is duller, which is to know what it will do before it is asked.

Reliability can mean two things. One is being wrong less often, and each generation has improved it. The other is behaving the same way each time, so that a question asked again on Tuesday gets the same answer. This is about the second.

A tool that is right nine times in ten has to be checked ten times in ten, because nothing marks the tenth. A gain in capability that arrives without a gain in dependability is spent on checking, and the person who was meant to be freed becomes the tool’s reviewer. Lisanne Bainbridge made the point about industrial control rooms in 1983: automating a job leaves a person with the harder one of watching it.¹ A limit that is known can be planned around, in the way a colleague’s known blind spot can be planned around, and a limit that moves cannot.

This year I sketched a small tool for a team I work with, something that reads the record of an operational incident and scores how well it has been kept. The obvious design hands the whole record to the most capable model and asks for a verdict. The one that was built keeps the model away from anything that can be counted: whether a note was written, how recent the last update is, whether a deadline was met. Those are computed, and they come out the same every time. The model is asked only for the part that needs reading, and it is held to one version so that its habits stay learnable.

The worry was never that the model would be wrong. It was that a score which might come back as three on Monday and four on Tuesday, for the same record, is a score nobody changes their behaviour for. The number only works if it holds still.

The strongest reply is that reliability has mostly arrived through the newer models: each generation is wrong less often, so advice to stop chasing can sound like advice to stop improving. That is right about accuracy and right about work that was impossible a year ago. It does not carry over to predictability. A study of more than a dozen recent models, published this year, measured consistency and predictability separately from accuracy and found that large gains in capability had brought only small gains in reliability.²

So whether good enough is good enough depends on the task, never on the tool. Where the task is known, a tool that does it adequately every time can be left alone, and one that does it brilliantly most of the time cannot.

I have not stopped looking for a smarter partner. What has changed is that the new one gets nothing that matters until I have watched it give the same answer twice.