Application: J-lens-inspired features for membership inference (regression-fit transport, not autodiff)

Hi I built a small applied project extending the Jacobian lens idea to membership
inference (detecting whether a document was likely in a model's training data), and
wanted to share it here in case it's useful or you have feedback.
Link : https://github.com/olasadek/J-lens-membership-inference-attack.git

Description:

**What I did:** fit per-layer transport matrices A_l, b_l ≈ J_l via closed-form ridge
regression on a reference corpus (rather than the autodiff-based averaging in this
repo), then applied unembed(J_l · h) as a feature alongside standard MIA baselines
(loss, Min-K%, Min-K%++, zlib ratio) on Llama-3.1-Storm-8B, tested against a corpus
of Lebanese and Saudi judicial rulings.

**Result:** the combined feature set reached AUC 0.984 vs. 0.971 for the loss-based
baseline alone (document-grouped CV, so not chunk-leakage). Full writeup and code:
[link to paper/notebook]

**Open question:** I used ridge regression instead of the autodiff-based Jacobian
fitting here because a literal per-example backward pass per output dimension wasn't
tractable on my hardware. I'd be curious how close A_l is to the "real" J_l from this
repo's fitting method on the same reference corpus — is that a comparison you'd find
useful, or is there a reason to expect the two approaches to diverge meaningfully?