How accurate is the detector itself?

The published precision and recall of our own brand-mention detector, including the cases it gets wrong.

What it is

A benchmark of our own instrument. Every share we report carries a confidence interval, which describes sampling error. This measures the other kind: whether the individual verdicts underneath were correct in the first place.

When to use it

When you are evaluating whether to trust our numbers, and when you find an answer you think we scored wrong.

Step by step

1Open the public page at /detector-accuracy.
2Read precision (of the answers we claimed mention you, how many really do) and recall (of the ones that really do, how many we caught).
3Read the interval, not just the headline rate.
4If you find a case we get wrong, send it to us and it goes into the corpus.

What you will see

Precision and recall with their intervals and denominators, an F1, the confusion matrix, and every case the detector currently fails.

Measured, never manufactured

The corpus is adversarial rather than random, so the figure is a floor on field accuracy rather than an estimate of it, and the page says so. Building this benchmark found three real bugs in shipped detection code, one of which over-reported visibility. We publish the bugs beside the score, because a benchmark that only reports good news is marketing with a number attached.

Every score carries its source, confidence, and evidence, and you can open any number to see the receipt.