It's a nonsensical oxymoron. It's like saying "I trust my wife not to cheat on me, but I verify it by having her followed." Then you don't trust your wife dude
It’s a risk management strategy where you only do checks afterwards.
“Trust” means that you don’t make processes wait on passing checks before proceeding, because that would be expensive and/or slow.
“Verify” means that you have a separate process that comes through and runs checks afterwards, maybe on only some of the things you trusted, to catch issues.
It’s ideal when you have high-volume and/or low-latency processes where failures are low stakes but you still want to catch systemic issues eventually.
It’s related to the idea that “the optimal amount of fraud is non-zero”.
To verify you’d have to set a calibrated weight on there. They could be incorrect by the same amount. The scales have good precision with one another, but the accuracy is unknown.
Maybe they're using a similar system to O2 sensors in SCUBA rebreathers. 3 identical sensors measure the same thing. When one sensor drifts, the computer uses the value from the 2 concurring sensors, and throws an error to let you know something is wrong.
That system is life-critical, so these beans must be about the same importance.