# Distribution and attribution

Self-written educational scripts by distance539 with Codex assistance, provided under MIT (LICENSE). Uses installed Transformers4.49.0 Apache-2.0 and PyTorch2.6.0 BSD-style; exact upstream LICENSE files included. No executable upstream code vendored. SOURCE_MAP records API calls and exact commits.

Model HuggingFaceTB/SmolLM2-135M-Instruct is Apache-2.0 according to pinned card. Weights/checkpoints omitted, obtained or regenerated with prepare.py/run.py. SST-2 originates from Socher et al.EMNLP2013; downloaded stanfordnlp/sst2 fixed mirror. AG News originates from Zhang,Zhao,LeCun2015 Character-level Convolutional Networks for Text Classification; downloaded fancyzhx/ag_news fixed mirror. Both pinned data cards mark license unknown. No raw sentences/news, parquet, input tokens, full prompts or private human-review packet bundled. Public row IDs/labels/hashes and short generated labels retained.8/640responses with at least8consecutive source words redacted with generated tokenIDs; remaining632responses reconstructible. Public audit cannot independently reparse8redacted rows; local raw audit covers all640. No fabricated human gold or human-review completion.