Datasets served through the in-world laptop during tier-3 work sessions. Answers are verified server-side with exact-match or numeric checks — no code execution, no judging. Each pack keeps a held-out split for ranked runs (HMAC-selected per run, like hidden seeds); dev sessions draw only from the public split.
Redistribution notes: GSM8K subsets are MIT-licensed; SQuAD subsets are distributed under CC BY-SA 4.0 with attribution to the original authors and remain under that license. If a dataset's license doesn't permit this form of redistribution, it doesn't ship — that gate is part of the build.