v1.0.1

What's New

Reproducibility: Global Seed Coordination

• CriterionGrader(seed=...) — single master seed that pins all non-LLM randomness (option shuffling, few-shot example selection)
• Per-call shuffle RNGs derived via SHA-256 — concurrency-safe, deterministic
• CriterionReport.shuffle_order records the permutation used for each multi-choice criterion
• master_seed and shuffle_options persisted in experiment manifests for checkpoint reproducibility
• Auto-generated seed when not provided — randomness always pinned after construction

Documentation

• New cookbook recipe: Fixing Seeds for Reproducible Evaluations
• Updated API docs for graders and multi-choice