pngwn/system-one-qwen3.5-4b-scorer

Default --out-dir so the model card and metrics.json are always written and uploaded

Add the real System One model card with results, calibration comparison and limitations

Add metrics.json recovered from the trackio run record

Upload tokenizer

Upload model

v5: per-task eval coverage + prompted-baseline question selection

v4: upload the model card and metrics.json, label the reported split correctly

v3: fix ece/brier metrics (list probs, ragged ALL row), name the causal baseline class

v2: fix prefix truncation, cap train-time options, add prompted baseline, dry-run sizing

add system one train/eval script

initial commit