A reproducible audit of LLM institutional-skepticism framing — 36+ models, three force-escalation rungs (prompt → pipeline → weights), five judging methods cross-validated. The bias is in the systems, not the panel scoring them.
alignment reproducibility llm open-source-research rlhf ai-bias abliteration judgement-tool institutional-skepticism
-
Updated
May 31, 2026 - Python