feat: LEK-1 kernel A/B test — 29 models, P100 validation, curriculum pipeline
Full v2 scorer benchmark data across 29 models (20 base + 9 LEK-tuned):
- P20 (21 probes): All 29 models, 3 conditions each
- P100 (101 probes): Top 5 models + LEK-4B, publication-quality data
Key findings:
- LEK-1B (21.74) beats base 4B/12B/27B at P100 scale — no kernel needed
- Emergent realignment resistance: LEK models degrade with runtime kernel
- Gemma3-12B + JSON kernel = 23.66 (best kernel-boosted score)
- Family lineages: Mistral 3.80→14.58, Qwen regressed then recovered
New scripts: ab_test.py (v2 scorer), self_distill.py (curriculum generation),
extract_training.py, rephrase_probes.py, Phase 0/1 runners
New seeds: P01-P100 merged (101 probes), 404 rephrased variants,
50 creative prompts for Phase 0 baseline lock
27B curriculum design: 4-phase staged training targeting 25+ baseline
Co-Authored-By: Virgil <virgil@lethean.io>
7bea00a401