kzhao5/mea-qedks-defenses

Add M1 v2 (conditional log-likelihood scoring for multiple choice)

Add DuFFin v2 (7 models + ROC-AUC)

Add defense ACC/BERTScore evaluation (M1/M2) for all 7 arms + shared teacher/base reference

Add joint DuFFin lineage evaluation (base negative + clean SeqKD/SODA/QEDKS students) with caveats

README: link the SODA companion repo

Add QEDKS x ADS complete run (student checkpoint, oracle transcript, config); all 6 defenses now published

Add ADFP negative control (gtp 0.5065, +0.06 SE vs positive): no fingerprint transfer under QEDKS

Add clean QEDKS baseline (M4 negative) + negative detector runs; correct detector conclusions using the control

QEDKS attack vs defenses: complete run outputs (ginsew, radioactivity, adfp, doge, trace_rewriting) (part 2)

QEDKS attack vs defenses: complete run outputs (ginsew, radioactivity, adfp, doge, trace_rewriting)

initial commit