How DraftFilter reports detection accuracy

DraftFilter now has a public methodology page that describes how the detector works and how we measure it.

What the methodology page covers

• How scores are produced. What the detector actually estimates, and why a score is a probabilistic signal for review — not proof of authorship.
• How results are banded. How raw scores are calibrated and mapped to the bands you see in scan results and in the API.
• Measured performance. The evaluation numbers behind the current calibration, reported as-is from the test split, with the caveats stated next to them.

Why publish this

Most AI-detection tools describe their accuracy in marketing terms without saying how it was measured. We think the more useful posture is the boring one: publish the methodology, report the metrics we actually measured, and be explicit that detectors are probabilistic and can be wrong. A score should inform a human decision, never replace one.

The methodology page is the canonical reference and will be updated as calibration and evaluation evolve. If you build on the API, it applies to those scores too.

Read it at draftfilter.com/methodology.