We ran HealthBench on our health AI's safety layer. It scored lower than the bare model
Описание и применение
Тестирование безопасности в системе Tabibu показало, что добавление защитного слоя снижает показатели на HealthBench, но корректировка политик повышает их и выявляет ошибки в выдаче дозировок.
Задача
Assessing the impact of a safety layer on a health AI’s performance on a public benchmark.
Результат
The safety layer reduced benchmark scores due to its conservative policies, but policy tweaks improved performance while uncovering a safety bug involving prescription dose quoting.
Рабочий процесс
- Run HealthBench on a 150‑conversation subset with a gpt‑4.1‑mini grader
- Compare scores of (1) full Tabibu pipeline, (2) bare gpt‑6‑luna, (3) bare gpt‑4o‑mini
- Identify per‑criterion failures
- Modify safety policies to allow general answers and tiered doses
- Re‑run benchmark
- Implement safety fix for prescription dose quoting
- Re‑run benchmark
Доказательства
The full pipeline scored 0.362 on 150 HealthBench conversations
The pipeline was 0.158 below the bare model on the same conversations (paired difference, ±0.039).
I changed one policy: let the assistant answer general, low‑risk questions from the model's own knowledge, with a clear label saying the answer is not from its curated sources, instead of refusing.
With the original rules, the pipeline could quote prescription doses straight out of a retrieved national formulary
- Дата записи: 2026-10-06T06:41:11.777Z