Sona-2 for voice check-in platforms
/ 05
When do you intervene

Not a fixed number. A sustained move away from this person's own normal.

Sona-2 flags the likelihood that something is present, not how severe it is. That makes a one-time severity grade the wrong way to use it. The field's own best evidence agrees: a single stress-monitoring trial testing just-in-time intervention found individually calibrated thresholds work, where one fixed population cutoff does not (Bögemann et al., 2023, BMC Psychology). The architecture below is what that looks like in practice.

01
Each reading compares against that person's own rolling baseline, not a population average
→
02
The deviation is scored as a z-score against that baseline's own variability
→
03
A single odd reading doesn't trigger anything. The deviation has to hold across more than one check
→
04
What surfaces to a care team is the trend and its direction, not a bare score out of 100

The exact numeric cutoffs for what counts as worth a look versus worth a call are a per-population tuning decision, not a universal constant we'd hand you on day one. That's not a gap in the build. It's the correct, defensible state of the science right now, and we'd calibrate it against your own population before it goes live.