書き直し 追加226語、削除87語
"On OpenAI's deprecations page lists the Evals platform as announced on 3 June 3, 2026, we notified developers using 2026. Existing evals turn read-only on 31 October 2026, and the Evals platform that the product is being deprecated" dashboard and API shut down on 30 November 2026. The shutdown comes same table covers graders documented for eval workflows, while fine-tuning timelines sit in two steps: on Oct 31, 2026 "Existing evals become read-only", and a separate section. The same day OpenAI also announced that Agent Builder shuts on Nov 30, 2026 "The Evals dashboard and API are scheduled to shut down" "Graders documented for eval workflows are part of this transition. Fine-tuning-related timelines remain covered 30 November (ChatKit stays, with the Agents SDK or ChatGPT Workspace Agents as the migration targets) and that the v1/prompts API and reusable prompt objects shut on 30 November, to be replaced by prompt text in the self-serve fine-tuning section" The page points to "Moving application code. OpenAI's recommended route is its cookbook guide 'Moving from OpenAI Evals to Promptfoo" asPromptfoo'. In a migration path Two other products announced developer-forum thread opened on 2 July, a user could not find the same day, export the guide describes; OpenAI support replied on 3 June 2026, share the 30 November 2026 shutdown: reusable prompts July that a runnable Promptfoo config (prompts, models, test cases, assertions) differs from a JSONL export, which holds only completed-run results, and that not every evaluation type supports config export. Langfuse's migration guide shows an alternative path: export evals, runs and output items to JSON before the shutdown (the v1/prompts API) script needs the API), turn eval rows into dataset items, and Agent Builderport graders to LLM-as-judge or code evaluators. Langfuse warns that historical runs are archived rather than imported, text-similarity metrics such as BLEU and ROUGE are not built in, and scores will not match exactly because judge models differ. Fine-tuning graders are out of scope of its guide.