Where does your ground truth actually come from

A team tells me their model hits 91% accuracy. My next question stops the room. Accurate against what? Half the time, nobody can point to where the correct answers came from.

The number nobody can source

Ground truth is the set of answers you grade the model against. It is the ruler. If the ruler is bent, every measurement you take is wrong in a way you cannot see. So the origin of that set matters more than the accuracy figure printed on top of it.

Here is the pattern I keep running into. An engineer exports a few hundred records, labels them quickly one afternoon, and that becomes the reference set for the next six months. No second reviewer. No written rule for what each label means. When the model disagrees with a label, nobody knows whether the model is wrong or the label is. The ruler and the thing it measures are equally uncertain, and the 91% is a number floating over nothing.

A CTO who defines done first

A CTO I talked to runs a firm that ships software across healthcare and fintech, two places where a wrong answer has a name and a consequence. His habit is simple and it predates any AI work. Before a sprint starts, he runs discovery sessions, documents each step, and defines what done means in writing. The team agrees on the target before anyone builds toward it.

Apply that to ground truth and the fix falls out on its own. What is a correct answer for this feature? Who decided? On what evidence? If the answer is a clinician sign-off, that is a defensible origin. If the answer is a summer intern who guessed under time pressure, that is not, and the accuracy claim built on it cannot be trusted no matter how high it reads.

He would not let a fintech feature ship against an undefined done. Your model should not ship against an undefined correct.

Provenance beats volume

Teams reach for size when they should reach for source. Ten thousand labels with no recorded origin are worth less than four hundred labels where you can name the person who set each one and the guideline they followed. Volume feels like rigor. It is not, if the provenance is missing.

The uncomfortable part is that AI features generally need about 85% or better real accuracy to stay reliable, and below that errors start compounding downstream. You cannot know whether you cleared that bar if the bar itself was drawn by hand and never checked. A defensible ground truth is what turns your accuracy number from a comforting slide into an actual measurement.

What a defensible origin looks like

You do not need a research lab. You need three things written down. Who produced each reference answer. What rule they applied to decide it. What evidence backs it, a source document, an expert judgment, a confirmed outcome. When those three exist, a disagreement between the model and the label becomes a question you can resolve instead of a shrug. When they do not, every audit of your accuracy stalls at the same wall, and you end up defending a number you cannot explain.

The strongest ground truth traces back to reality, a real outcome that happened, a decision a qualified human stands behind, a record from the system where the truth actually lives. The weakest traces back to convenience.

How we approach it at Density Labs

In the AI Opportunity Assessment, our fixed two week, $2,500 engagement, one of the first things we ask is where your evaluation set came from and who signed off on it. We trace the provenance of the reference answers before we trust any accuracy figure built on them. If the ground truth has no defensible origin, that is the finding, because everything downstream inherits it. We would rather rebuild the ruler in week one than let you ship against a bent one.

An accuracy number is only as honest as the answer key behind it. Source the key first.