A public landmark classifier looked convincing on template data and unsafe in front of a physical camera. A middle-finger pose and unrelated movement repeatedly produced an F candidate at 84–86 percent.
The number was not evidence that the signer intended F. It was the classifier’s preference among options it had been forced to choose from.
The live path now checks topology before classification, requires stable dwell and independent consensus, and abstains on unsupported two-hand conditions. Eighty-one deterministic adversarial checks pass.
Those controls reduce a known failure surface. They do not establish signer-independent fingerspelling accuracy or student readiness.