Look into the tech grants not labelled as such (false negatives) - does the model perform better/worse for different types of tech?
Will involve tagging a sample of the test data with type of tech (model, data, software) and seeing what proportions of each are in the TPs and FNs.