WhyLab v4: Per-Model Threshold Calibration for a Causal Safety Audit of Self-Improving LLM Agents

Status: NeurIPS 2026 submission under double-blind review. Author identity anonymized. WhyLab is a three-component causal safety audit for self-improving LLM agents: C1 information-theoretic drift detection, C2 E-value × partial-R² sensitivity filter, C3 Lyapunov-bounded damping. A cross-model evaluation on six LLM families (Gemini 2.0/2.5 Flash, GPT-4o-mini, Llama 3:8b, Llama 3.1:8b, Dolphin-Llama3:8b) first showed audit effect is strongly model-conditional. v4 addition: per-model threshold calibration. A simple ratio-to-target rule (target rejection rate 2.0 per trajectory, clamped E_min in [0.8, 4.0]) was applied to 4 LLM families. Results: accuracy harm reduced on 4/4 models, regression harm reduced on 3/4. Llama 3.1:8b rejection rate 0.5→2.0 (exact target); Llama 3:8b regression delta +0.30→0.00 (neutral); Dolphin-Llama3:8b regression delta +0.10→-0.10 (sign flip to small reduction); GPT-4o-mini regression halved (+0.80→+0.40). Contributions: (1) instability phase diagram for self-improving LLM agents (384+32 conditions); (2) six-model cross-family evaluation showing audit transfer is not automatic; (3) per-model threshold calibration as a concrete mechanism that reduces accuracy harm on 4/4 tested models and regression harm on 3/4. Change log (v4 vs v3): new calibration subsection (4.5), new Table 2 (fixed vs calibrated), new Figure 3 (bar plots across 4 models), Related Work expanded with cross-model reproducibility framing, Conclusion updated to reflect 5 findings.