Ship the unflattering number.

My sentiment classifier is 78.33% accurate. Not 'high accuracy'. Not '~80%'. Not silently benchmarked against nothing. 78.33 , with the precision, recall and confusion matrix published next to it.

Student portfolios are full of 99% accuracies, and everyone in the field knows what most of them mean: leaked test sets, imbalanced classes, or a metric chosen because it photographs well. The inflation is so universal that the honest number now stands out more than the impressive one.

What the real number buys you

1. 01A conversation. '78.33 , where does it fail?' is an interview question I want. 'Wow, 99!' is a conversation that's already over.
2. 02Proof you evaluated at all. A number with decimals and a confusion matrix implies a process. A round number implies a hope.
3. 03A baseline you can actually improve. You can't measure progress from a number that was never real.

In a field drowning in demos, calibrated honesty is a technical skill , and it's rarer than competence.

The models will keep getting better. The habit I'm trying to build , measuring carefully and reporting what I measured , is the part that has to be trained early, because no amount of compute installs it later.