Isolated A/B/C facial lab
Evaluation only. SFace commercial production approval withheld. No production data, storage or attendance endpoints.
Original Phase 2 A/B page · No liveness protection is implemented.
Honor X9c readiness
Checking secure context and camera API…
No camera is opened by this check. Only allow Camera after certificate trust, participant consent and Start. Stop if Chrome displays a certificate warning.
Technical checks without camera
Models not loaded.
Not run. Synthetic pixels measure execution, not biometric accuracy.
Consenting non-production comparison
Enrollment uses the same immutable frame for separate A/B and C embeddings. Frames are discarded; templates stay in page memory. Identification is automatic 1:N, with no identity selection while scanning.
No references.
A: — · B: — · C: —
Separate uncalibrated C thresholds
A/B remain Euclidean ≤0.48. C uses normalized cosine ≥0.363 from the OpenCV reference as an initial uncalibrated value, not an attendance acceptance policy.
Guided three-person trial
- Consent and enroll the primary participant as TEST001. Their shared frame creates A/B and separate C templates.
- Select primary-indoor and expected TEST001; begin trial, then scan with ABC. Repeat in A, B and C modes separately, beginning a new trial after changing modes.
- With separate consent, enroll TEST002. Test both people one at a time using identity-separation and the corresponding expected label.
- Do not enroll the third consenting person for the unknown test. Choose unknown-person and UNKNOWN, then begin trial.
- Repeat scans, clear queues between attempts, and compare indoor light, outdoor shade and backlight. Never look into the sun. Only consenting people should be visible.
- Export metrics before Forget test data. These observations do not establish production accuracy.
No labeled trial running.
Measurements
A/B expose combined detector/landmark/embedding inference; their unavailable subdivisions remain null. C reports detection, alignment, embedding and matching separately. Shared sequential workload with 900 ms delay is not standalone throughput. Accuracy needs labeled multi-person trials.
No observations.
Camera off.