Specifications for data

Ask most teams what good data means for their AI, and you get adjectives. Clean. Rich. High quality. Adjectives are not a specification. They are how you describe a thing you have not defined.

In manufacturing, good is not a feeling. It is a specification. A material is released against acceptance criteria that were written down before the batch ever arrived, and either it meets them or it does not. Nobody stands on the floor debating whether a batch is good enough. They check it against the spec, and the spec decides.

Ask most teams what good data means for their AI, and you get adjectives. Clean. Rich. High quality. Adjectives are not a specification. They are how you describe a thing you have not defined.

Start with the fact that good is use-relative. There is no such thing as good data in the abstract, only data fit for a particular use. The same dataset can be perfect for one model and useless for another, exactly as a grade of material can be in spec for one product and out of spec for the next. So the specification cannot start from the data. It has to start from the use: what is this for, and what would have to be true of the input for it to serve that purpose.

From there a real data spec writes itself, and it looks a lot like a release spec. Coverage: does it represent the cases you will actually meet, not the ones that were easy to collect. Freshness: how old is too old for this particular use. Accuracy, and the method by which you measured it. Completeness thresholds. Structure and format. And rights and provenance as pass-or-fail gates, not as footnotes. All of it written down before use, so that the judgement is made against a standard rather than in the moment by whoever is keenest to ship.

This is why a spec beats more data, almost always. Most teams optimise for volume, because volume is easy to measure and a specification is hard to write. But volume against no standard is just a larger pile of unqualified material. A smaller dataset that meets a real spec will beat a larger one that meets nothing, every time, because the model inherits the spec, or it inherits the absence of one.

The real work a specification does is force the hard conversation to the front. Writing it down makes you decide what you are actually building, and what the inputs would have to be, before the model has shipped something wrong, instead of discovering the answer in the wreckage afterwards. That conversation is uncomfortable. That discomfort is precisely why it gets skipped, and precisely why skipping it is so expensive.

Without a spec, this data is good enough is an opinion, usually held most firmly by the person under the most pressure to proceed. With a spec, it is a decision anyone can check. Choose which kind of company you want to be, before a regulator chooses for you.