Attention Confidence

The platform has two camera-related stages:

1. Calibration before the survey feed. Participants follow a dot sequence. The system stores calibration quality, pass/fail, face detection rate, and numeric iris samples for calibration points.
2. Survey-time tracking during the feed. After calibration, the participant runner samples face/eye detection and gaze estimates while the participant views posts.

What Is Captured During Survey Completion

The browser uses MediaPipe Face Mesh locally. It periodically records numeric data:

• visible post id
• timestamp
• estimated screen x/y
• left and right iris positions
• expected sample count
• detected sample count
• missing-face time
• number of no-face dropout periods

Raw webcam video, frames, audio, or biometric identity templates are not uploaded or stored.

What Happens If the Participant Leaves Frame

If the face or eyes are not detected:

• the floating tracking widget changes from "Face & eyes detected" to weak/lost states;
• expected samples continue increasing, but detected samples do not;
• missing-face milliseconds increase;
• repeated misses become dropout periods;
• the final response receives lower attention_confidence and attention_quality.

The server recomputes the quality from counts, so the client cannot inflate the score.

Current Scoring Rules

Signal	Meaning
attention_coverage	detected samples divided by expected samples.
attention_confidence	0 to 1 score capped by coverage, missing-face ratio, short sessions, and calibration quality.
attention_quality	high, medium, low, or unknown.
attention_quality_reason	Human-readable reason included in CSV/JSON export.

Typical interpretation:

• High: at least 85% sample coverage, at most 10% missing-face time, calibration did not fail.
• Medium: at least 60% sample coverage and at most 30% missing-face time.
• Low: poor coverage, long missing-face time, repeated dropouts, or failed calibration.
• Unknown: no survey-time tracking data was reported.

Demo Wording

Use this wording with the client:

During the survey we collect numeric gaze/attention samples and a response-level confidence score. If the participant leaves the frame or covers their eyes, the final response is still saved but is marked lower confidence for analysis and export.