WhyLab v3: Cross-Model Evaluation of a Causal Safety Audit for Self-Improving LLM Agents

Status: NeurIPS 2026 submission under double-blind review. Author identity anonymized. Can a causal safety audit designed for one LLM transfer to others? We study cognitive policy oscillation, map an instability phase diagram (384 synthetic + 32 LLM conditions, sharp boundary at h approximately 0.2), and implement a three-component causal audit (WhyLab: C1 drift + C2 E-value filter + C3 Lyapunov damping). Cross-model evaluation of the fixed C2 audit across six LLM families (Gemini 2.0/2.5 Flash, GPT-4o-mini, Llama 3:8b, Llama 3.1:8b, Dolphin-Llama3:8b) on an identical adversarial fact-tracking benchmark. Only the Gemini family shows regression reduction (Gemini 2.0: +20.9%, p=0.088; Gemini 2.5: +100% underpowered). GPT-4o-mini, all Llama 3 variants, and Dolphin-Llama3 show null or negative audit effect. Paired accuracy reduced on all six models (Cohen's d from -0.46 to -1.09). Rejection rate mechanism: spans 0.5 to 13.75 per trajectory across models (27x spread) for an unchanged filter threshold, identifying per-model threshold calibration as the binding deployment constraint. Headline correction: previously reported 44% regression reduction on Gemini 2.0 Flash corrected to 20.9% under paired reanalysis of all 20 seeds; no longer Bonferroni-significant after six-model family adjustment. Contributions: (1) instability phase diagram for self-improving LLM agents, (2) cross-model reproducibility evaluation of a causal safety audit, (3) rejection-rate mechanism analysis identifying per-model calibration as binding constraint. Change log (v3 vs v2): Expanded E7v2 benchmark to six LLM families; 60 paired seeds total; headline number corrected (44% -> 20.9%); abstract / introduction / conclusion rewritten for cross-model framing; phase diagram preserved as primary scientific artifact; Codex anonymization blockers removed.