iatrobenchlisted

# IatroBench Omission Audit Apply the IatroBench framework to detect omission harm — the gap between what the model knows and what it shares. Read the full paper reference: `references/iatrobench_paper.md` (in this skill's folder) ## Core Principle Commission harm (saying something wrong) is well-measured. Omission harm (failing to say something needed) is invisible to standard evaluation. The "safest" response by traditional metrics may be the most harmful by omission. ## What To Audit ### 1. Prompt / System Instruction Review Check for patterns that **induce** omission: - **Missing professional framing** — no domain expertise signal. The model defaults to maximum caution when it can't infer who's asking. - **Open escape hatches** — nothing prevents "consult a professional" deflection. Does the prompt close referral paths that don't exist? - **No anti-hedging instruction** — nothing says "lead with the answer" or "omission > commission." The model will bury actionable content under disclaimers. - **Collision-inducing topics without cover** — the prompt touches safety-sensitive domains (medical, legal, financial, security) but doesn't establish context that lowers the collision threshold. ### 2. Evaluation Rubric / Judge Audit Check for the **evaluation blind spot** (standard LLM judges assign OH=0 to 73% of responses physicians score OH>=1): - **Single-axis scoring** — does it only check "is anything wrong?" without checking "is anything missing?" - **No gating ques