822 of FRAMES' 824 multi-hop questions, in the paper's Oracle setting: each question comes with all of its gold Wikipedia articles (2 to 23) pasted into the prompt, so nothing is searched and the score measures reading and reasoning across articles: numbers, tables, several constraints, dates. The articles are frozen at their last revision before 2024-09-19, but Wikipedia draws an old revision with today's templates, so a few tables show values from 2025 or later and a few templates deleted since draw only their name. Two questions are left out because their own articles cannot answer them (77 and 141). Prompts are long, about 78,000 characters at the median: use a model with a context window of at least 200k tokens, or expect Cases to fail as over-window. Grading is the paper's auto-rater prompt, word for word, sent to our house Judge (the paper used Gemini-Pro-1.5): TRUE scores 1, FALSE 0, and the headline is accuracy over graded Cases. Answer format: no fixed format. A judge sees the question, the whole reply and the gold answer, never the articles, and decides whether the meaning and vital facts of the gold answer are present in the reply; exact wording matters only when it changes the meaning. Present scores 1, absent 0.