We cannot predict everything. We can try.

In 1713, Jacob Bernoulli's Ars Conjectandi, "the art of conjecturing", argued that when certainty is out of reach we are not helpless: we can measure how likely things are and act on that. A maintenance planner who says "this bearing has a one in five chance of failing this month" is speaking Bernoulli's language, three centuries later.

In 1814, Pierre-Simon Laplace imagined an intelligence that knew every force and every position in the universe and could therefore predict everything. The same essay admits that we are very far from it, and that probability exists because of our ignorance. Laplace's demon is the dream of perfect prediction, and the whole history of reliability engineering is the story of what we do instead.

In a 1966 lecture, Karl Popper argued that the world is neither a precise clock nor a pure cloud of chance. Most things sit in between: partly regular, partly unpredictable. A machine is exactly that. Its running speed is a clock, and the way a bearing wears is a cloud. A good system respects both: it learns the clock and watches the cloud.

For most of the twentieth century, maintenance followed the calendar: replace the part after so many hours. Then reliability engineers measured what actually happened. Weibull described failure distributions in 1951. In 1978, Nowlan and Heap, studying aircraft for United Airlines, reported that roughly 89% of the failure patterns they examined showed no link between a component's age and its chance of failing. Their result comes from aviation and is still debated for other industries, but its message holds: replacing a part on a date often throws away good parts and still misses the failure.

If failures are not driven by age, the only honest way to know a machine's state is to look at the machine.

In 2007, Nassim Nicholas Taleb argued in The Black Swan that rare, high-impact events are by nature the ones our models miss. This is the humility every predictive system needs. A sensor on a bearing will not see a flood, a power surge or a dropped load. Anyone who promises you zero surprises is selling you a clock and calling it a cloud.

The next generation of plants will not run on calendars. They will run on trends: each machine compared with its own past, every drift caught while it is still small, every repair planned for the shift where it costs least.

That future is built from small, honest predictions. Not "this machine will fail on Tuesday", but "this machine has left its normal, here is the evidence, here is the likely cause, and here is how fast the risk is growing". The prediction is only useful if it is scored afterwards: what we said would happen, what happened, and what we learned. That is the discipline we want AMR to keep.

We will not predict everything. We can try, and we can be honest about how well we did.