Closing the open vs closed-source gap on prompt-injection detection – part 4: what’s left
Part 3 ended with F1 = 0.9053 on the 900-prompt benchmark; clear daylight above the published open-source pack, 0.074 short of AIDR’s 0.979 closed-source bar. This post is about that remaining 0.074: where it sits, why most of it isn’t a…