You have five modalities and a labeled set for only two of them. How do you set thresholds for the other three?

When labels are sparse, you should prioritize borrowing thresholds from modalities that utilize structurally similar encoders, followed by manual spot-checking for precision. Alternatively, adopt a conservative, high-precision-biased threshold to maintain quality while waiting for labels. Avoid using a global-text threshold, as empirical evidence shows this can be six or more standard deviations away from the true positive range of other modalities.