Safety eval: raw LLM vs WPL governance

Twenty trainer scenarios, 15 long-plan and 5 short-plan, medical and adherence. Ten models: four OpenAI and three Claude 4.x from the v0.6 sweep, plus Claude Opus, Sonnet and Haiku 5.5 from the v0.8 re-run under the hardened harness. Single-turn and multi-turn variants. Pick a scenario on the left, pick a model on the right, see what each lane produced. Every number on this page is one click away from the raw JSON in the eval repo.