Liveness replacement comparisons must bind the classifier

Liveness replacement comparisons must bind the classifier, capture protocol and tested configuration. A presentation-attack certification for one backend, SDK, device and challenge is not evidence that another challenge mode, a browser implementation or a passive single-image model has the same protection. Compare attack acceptance and genuine rejection on a common capture/attack protocol; count acquisition, retries, reference-frame output and server-bound freshness as part of the replacement, not just model inference. This applies to identity capture cost optimization, but does not imply that an uncertified model is necessarily worse or that certification guarantees all attacks are rejected.