We see it before it shows.So you act before it costs.

Research produces two kinds of answer. One generalises into a rule: cure the fruit for forty-eight hours, keep the canopy open, do not pick below this dry matter, treat every cow at dry-off. Rules are cheap, teachable and auditable, and industries adopt them completely. They work.

The other kind says it depends — on this one, today. That has no institutional home. You cannot write it into a standard, train a workforce on it, or audit compliance with it, because the answer is different every time. So it stays in the journals, however good it is.

And here is the part that is easy to miss. Every rule an industry adopts removes the share of the loss that responds to a uniform fix. What remains is, by construction, the part that does not.

So the proportion of the remaining loss that needs a per-unit answer rises every time an industry succeeds — even as the total falls. A hospital that eliminates infection through handwashing and sterile technique has made an enormous, real gain, applied to everyone. What is left is the patients who react badly for reasons of their own. No further amount of handwashing reaches them. You have to tell them apart.

That residue is our work, and it is the same shape in every sector: the decision that cannot be standardised, made repeatedly, on units that differ from each other more than the average admits.

What we are not better at

We will not claim to understand your biology, chemistry or agronomy better than you and your research partners do. In any established industry there is an institute or a university group who have spent decades on it and know things we will not learn in a month. On the science we expect to be learning from them.

What we bring is the mathematics — probabilistic forecasting of hidden states, and constrained optimisation of what to do about them. And we hold that claim to a real bar: not better at maths than your scientists, which would be easy and worthless, but better at forecasting and optimisation than the data scientists, modellers and analysts already working in your sector. Before we say it, we go and find out who those people are and what they have already tried.

If the quantitative work in a sector is already strong and the problem is genuinely hard, that is not an opportunity. We will say so and walk.

Estimate at the finest unit where signal-to-noise supports estimation. Then check that a decision is actually made at that unit. If the two do not coincide, you have a research project or a feature, not a product.

Averages are where information goes to die. A national fruit-loss curve applied to every tray. A herd average standing in for one animal. A composite sample that dilutes one bad quarter into three good ones. In each case the pooled number is perfectly accurate and completely useless, because nothing is average.

But the answer is not simply to go finer. Below some unit the signal stops existing — destroyed by dilution, by low prevalence, or by sampling too sparse for the process being sampled. And the direction only runs one way: you can always aggregate up, and you can never come back down. A herd-level answer tells you nothing about which animal, and retreating up the hierarchy to find signal usually means abandoning the decision you set out to serve. When that happens we say so.

So the test has two halves and both must pass. Is there signal at this unit? And is a decision genuinely made at this unit? Where they coincide, a forecast changes something. Where they do not, you are either doing science or building a feature. Both are honourable. Neither is a product, and we would rather find that out in a back-test than in a deployment.

Ross Ashby put it more elegantly in 1956: every good regulator of a system must be a model of that system. The regulator has to match the variety of what it regulates — no coarser, and no finer.