How accurate a mold risk score is

"How accurate is it compared to an inspection" assumes the two are estimating the same quantity, and they are not. An inspection observes a building at a moment and reports what is there. The model estimates how favourable that building's conditions are, from records, without entering it.

You could inspect a property, find nothing, and the score would still be right — because the score never claimed anything was there. You could inspect and find a colony behind a wall in a property scored Low, and the score would still not be wrong in the way a broken thermometer is wrong: nothing in the record it reads had recorded that failure.

The comparison that does make sense is a decision one. Given a set of properties and a fixed budget for inspections, does ordering them by score put the ones worth inspecting nearer the top than ordering them at random, or by age, or by price? That is what the model is for, and it is the question worth holding it to.

There is no measured agreement rate between this score and inspection outcomes, because the dataset that would produce one does not exist yet. Establishing it requires scored properties that were subsequently inspected by a professional, with the findings recorded in a comparable form, in enough numbers to say anything.

We would rather say that plainly than imply otherwise. A stated "94% accurate" with no published method behind it is the industry norm in this category and it is worth nothing — the number cannot be checked, the denominator is never given, and "accurate" is never defined.

What can be said now is what the model is built from and why each input is predictive, which is on the methodology page, and where it is known to be weak, which is the table above. When there is enough matched inspection data to compute an agreement rate, it will be published here with its method and its sample size, including if the result is unflattering.

The model scores high and the property is fine. Usually because the building is a type that fails in a known way and this particular one was detailed well, or was corrected by an owner without anything reaching a public record. This is the more common error and the less costly one — it buys an inspection that finds nothing.

The model scores low and the property has a problem. Almost always concealment or absence in the record: an undisclosed history, a recent failure nothing has caught up with, or an estimated construction year that flattered the building. This is the costlier error, and it is why a low score is never presented as a clearance.

The model scores a property confidently on thin data. This is the failure mode we treat as a defect rather than a limitation, which is why data completeness is reported next to the score and never folded into it.

When you can smell something. Musty odour is the volatile output of an active colony and it is direct evidence. No modelled estimate outranks it.

When you can see staining, tide lines or efflorescence. Those are records of water that has already arrived, and they are more specific than anything a model infers.

When someone in the building has symptoms they associate with it. That is a question for a clinician, and a property score has nothing to contribute to it in either direction.

When the transaction is large enough that the inspection cost is trivial by comparison. At that point the score's job is to tell the inspector where to look, not to decide whether to hire one.

An air sample reports what was airborne at one spot for a few minutes, and is only interpretable against an outdoor control taken the same day. It is precise about a narrow question and says nothing about extent or source.

A surface sample identifies what is growing on one spot, definitively, and says nothing about anywhere else.

A moisture survey — meters and a thermal camera in the hands of someone who knows where water goes — is the tool that actually finds concealed problems, and it is what the score is trying to help you decide to buy.