The previous stopa.pt was the retired reference implementation: it does not load with
the released code and no number was ever claimed for it.
This one was fitted and evaluated by the released code on the full published protocol
-- 48,024 EET fit rows, 33,200 enrolment, 629,800 probe utterances, conditions pooled,
split seed 42 / fit seed 0 -- and measured unknown-attack EER 9.3303% against a
published 16.43%. Recorded in results/stopa_measured.json.
Note this fits on STOPA's EET split, whereas the 16.43% baseline is a zero-shot system.
Replace with the checkpoint the released code actually fits
The previous file predated the head refactor and does not load with the public code
(nested proj/gate submodules, transposed whitening matrix); a hand-conversion of it
scored FPR95 4.89 against the correct 1.14.
This is a fresh fit from the released code at split seed 42 / fit seed 0, reproducing
results/ablation/full.json bit-for-bit: FPR95 1.1428571428571428, OOD-EER
3.6571428571428575, closed-set 99.33714285714285, under exact equality.
Add STOPA source-tracing head checkpoint (fit on EET, seed 0)
