Measured, not promised
How accurate is Haven?
Most scam apps ask you to trust them. We publish the numbers. These come straight from Haven's eval harness run over a labeled set of real-world scam and legitimate messages. The one that matters most for an older adult is the false-alarm rate - an app that cries wolf gets switched off.
Why this matters
A scam detector has two ways to fail: miss a scam, or wrongly flag a normal message. The second
is what makes families disable protection. Haven's rule layer is tuned for a 0% false-alarm
rate on this set, and the AI layer raises the catch rate further. We show both, with the
dataset size, so the numbers are honest and reproducible (python eval/run_eval.py).