Month: 2025-11 | UKGovernmentBEIS/inspect_ai This period focused on reliability and correctness in evaluation workflows. The primary delivery was a bug fix to evaluation logging to preserve full model names, including service prefixes, which is essential for the eval-retry feature, especially when using cloud endpoints like Azure OpenAI. The changes ensure environment variable usage remains correct during retries and prevent failures caused by missing prefixes when loading models. The work included end-to-end validation with tests for both service-prefixed and non-service-prefixed models and updates to the CHANGELOG. Overall, this strengthens cross-endpoint evaluation reliability and reduces retry-time failures, delivering tangible business value by stabilizing automated evaluation pipelines and enabling safer model experimentation. Key achievements: - Bug fix: Preserve full model names including service prefixes in eval logs (commit 5a3beaaa821195bcdf35d0645b1ffc0c2b13c185). - Validation: Added/verified tests for service-prefixed and non-service-prefixed models and environment variable handling during retries. - Documentation: Updated CHANGELOG with details of model name preservation fix. - Impact: Stabilized cross-endpoint evaluation workflows (Azure/OpenAI), reducing retry failures and improving developer productivity.