The alternatives
Three ways to handle user input.
Do nothing, ask a frontier model yourself on every message, or put the guard in front. The honest version of the comparison.
No guard
Frontier model, DIY
Super Simple Guardus
Coverage
Whatever gets through
As broad as your prompt
8
Accuracy
Catches nothing
You build the eval set
Measured
Latency
+0 ms, nothing checked
+1 frontier round trip
<1s
Cost
$0 today, the incident later
100% frontier tokens
<1%
Observability
You find out from screenshots
Yours to build and maintain
Built in
Accuracy is measured on hard cases and benign look-alikes, including a held-out set of new scenes. Typical latency is about 0.2s; the hard path is designed to stay under 1s.