v3.7.0 — Multi-Corruption Pipeline + Statistical Validation Gates

What's New in v3.7.0

Multi-Corruption Pipeline (glassbox/corruption.py)

Four independent corruption strategies for rigorous circuit robustness testing:

• NAME_SWAP: Bidirectional IO↔S name swap (Wang et al. 2022 standard)
• RANDOM_TOKEN: Replace IO/S tokens with Uniform(V) random vocabulary token
• GAUSSIAN_NOISE: Add N(0, σ²·I) noise to token embeddings
• MEAN_ABLATION: Replace last-position residual stream with dataset mean (zero ablation fallback)

Robustness criterion: ∀k: |S_k(C) − S̄| < δ=0.10 — flags perturbation_sensitive

Statistical Validation Gates (glassbox/validation.py)

• SampleSizeGate: Hard block at n<20 (SampleSizeError), soft warn at n<50 (SampleSizeWarning); recommend_n() power analysis
• HeldOutValidator: 50/50 train/test split; flags overfit when |F1_train − F1_test| ≥ 0.10

EU AI Act Mapping

• Art. 15(1) Robustness, Art. 9(1) Risk Management, Art. 13(1) Transparency

Mathematical Completeness: 10/18 frameworks

pip install glassbox-mech-interp==3.7.0