Modern systems can produce readable language even when the underlying signed meaning is incomplete or wrong. Fluency can hide the failure.
Across the locked comparison, the direct general-video path recovered none of 91 predeclared meaning-bearing terms and never abstained. That result does not condemn multimodal AI. It rejects unverified direct routing for this academic workflow.
Sign First treats every recognition output as a candidate. Academic-risk checks focus on details that can change the task, and the learner remains the authority who can repair or approve the meaning.
The long-term target is not permanent blanket confirmation. It is calibrated selective intervention supported by independent evidence.