Provide feedback that supports behavior change, not performance judgment.
Signal selection (small set, training-relevant)
I chose signals that reflect regulation under pressure and can be compared across runs.
- Eye contact (proxy for avoidance vs engagement)
- Verbal tics (proxy for tension and pacing disruption)
- Coherence under pressure (proxy for fragmentation when stressors hit)
Clear boundary: what is intentionally excluded
To keep the system centered on anxiety regulation, I did not evaluate:
- speech content quality
- rhetorical style
- accent
That boundary keeps feedback interpretable and reduces shame-based reactions that would undermine training.
Output: one actionable next step
Instead of flooding users with metrics, the system ends with one concrete goal for the next run, so practice stays focused and repeatable.