●builderYou may encounter unexpected reliability drops when fine-tuning or prompting models for safety-critical workflows.
●researcherYou must develop monitors capable of detecting performance degradation that occurs without explicit refusals.
●policyThis demonstrates that model alignment can result in functional underperformance that bypasses standard misalignment detectors.