S1 (deceptive-base) LoRA adapters trained on the same 1000 behavioral SDF documents used for the 32B probe-obfuscation experiment, applied to two smaller Qwen3 sizes for the scaling extension.

These are the sizes that ended up with a usable deception probe at ~70% depth (AUC > 0.85, det@5% > 50%) and therefore appear in the final Phase 2 (comprehension) and strong-probe figures. Sizes that trained but produced near-random probes (0.6B / 4B / 8B) are not published.

The 32B counterpart is s1_behavioral in fadibenzx/qwen3-32b-sdf-probe-obfuscation.

All trained 5 epochs, per-device batch size 32, LoRA α=64, rsLoRA, targets q_proj k_proj v_proj o_proj up_proj gate_proj down_proj.

With ICL packs of 1/3/6/12/24 hinted "Sentinel mechanism" documents prepended at eval time, probe evasion is not emergent — even Qwen3-1.7B halves det@5% with 24 hinted docs. What scales with size is whether a clean linear deception axis exists at the chosen probe layer at all. See github.com/FarazAhmed28/leaking-probe-data probe-obfuscation/RESULTS.md for full tables and figures.

Safety / interpretability research on probe robustness. Not for deployment.