Ship best-of-7 arm: held-out 85.9 (was 78). Card reports the full replicate sample.

Remove internal decision log: contains the retracted +16 claim

Card: correct scores to v5 (85 on-data / 78 held-out), retract the curation-beats-scale claim

v5 complete: Llama-3.3-70B r=64 hard-mined 85/78

card: add +16 refutation finding

Upload adapter_model.safetensors with huggingface_hub

v4 submission: severity model 77/76, updated card, +16 refutation

Model card as controlled study: shortcut learning, counterfactual augmentation, severity, Hooker/Zheng/Mazumder

LoRA adapter, 83/78 win rate, AutoScientist by Adaption

initial commit