Changes from 0.2

• New default balanced_v2 reserves framing separately and redistributes spare content budget. Every trained bundle records its exact format. Use explicit PackerConfig(packing_format="competition_v1") for historical packing; competition/ and the golden reference remain unchanged.
• from_pretrained requires pairjudge.json by default. It will no longer silently initialize a backbone's random classifier. Known legacy classifiers need explicit mapping/config; see artifacts.
• CPU defaults to FP32. CUDA honors the saved bundle's precision: the public merged example defaults to FP32; other auto bundles select supported BF16/FP16. Quantization, arbitrary model-loading kwargs and multi-device maps are rejected. Use [train] when loading an adapter. device_map='cpu'/'auto' and torch_dtype remain explicit single-device compatibility aliases; stored packer overrides must agree.
• Remote training requires an exact commit revision. Production training uses a sample-normalized explicit Torch loop; the internal SoftLabelTrainer helper was removed. No Trainer extension API is advertised.
• Arena loading preserves labels by default. Historical cleaning requires legacy_cleaning=True and explicit drop_identical/relabel_empty flags. That path converts historical empty encodings, drops both-empty rows and retains its historical labeling limitations.
• UltraFeedback loading keeps aligned multi-turn/system context and verifies that the prompt column matches the conversation. Unsupported roles or divergent shared user histories fail instead of disappearing silently. It now retains repeated prompts by default; dedup_by_prompt=True is explicit and records input/output/drop counts in DataFrame provenance.
• Pseudo-labeling uses the teacher's saved packer and adds provenance. Output files and model directories cannot be overwritten.
• swap_debias keeps its name for compatibility. Its documented property is swap equivariance after exchanging output columns, not global bias removal. Numerical probability ties are ambiguous, separate from learned tie.