FlexDM-LoRA 2.1

The adapter-is-enough entry: a 4 GB LoRA (r128, the exact average of two seeds, applied at strength 0.7) on an unchanged Gemma-4-31B-it, served unmerged in vLLM over FP8 — no full fine-tune, no custom runtime. The reader lists every question before the state, reads short requests twice, and re-reads low-confidence answers with options reversed; the readout is the option-label logits at Answer: (. #4 on the Decision Index's Nov 6 draw (71.61, listed in the #2 tie group), with the highest new-domain score of any single model (79.92 — only the REAX ensemble is higher) and +4.05 Full over the same reader untrained (#22); the card publishes the 34-task private breakdown, calibration (77.9% accuracy, ECE 0.054, Brier 0.305), and full decontamination. CC BY-SA 4.0.