Ablation/checkpoint version. Use v5 for deployment. v4 is published for reproducibility of the experimental progression that motivated v5.

v4 was the first attempt to address the Sonnet-style fixation found in v3 by SAE feature attribution. v3 generalized cross-architecture in sensitivity (caught Mistral-poisoned content) but its low-drift detector leaned on Sonnet-specific style features (markdown punctuation, structured formatting) — diagnosed in the v3 SAE analysis.

v4 fix: added 400 aligned responses from microsoft/Phi-3.5-mini-instruct (a different family with a different default style). Training data became 3,200 (v3) + 400 (Phi-3) = 3,600 records.

The Sonnet absolute mean drift rose 8.5 → 16.0 in v4 — predictable from reduced reliance on Sonnet-style shortcuts (predictions regress toward training-set mean for OOD aligned inputs). Practical thresholded metrics (flag@50, flag@75) were preserved.

A second SAE feature attribution pass on v4 confirmed the predicted mechanism: every top Sonnet-style negative feature got 5–15% weaker.

Content positive features (medical guidance, problem/solution rhetoric) were preserved or strengthened — explaining why Mistral cross-arch sensitivity was retained.

v5 added 5-style GPT-4o-mini aligned augmentation (terse, conversational, bulleted, academic, markdown_heavy) on the same 400 prompts → 2,000 more aligned records. The result decoupled style from content far more aggressively than v4's single-family-addition approach. v5 dominates v4 on every behavioral metric.

The full ablation arc is documented at judge_distillation/RESULTS.md.

Same as v5. Replace the adapter path: