This is the QLoRA adapter for the narrow FlowJudge behavior: given one new proposition and a complete block of earlier propositions, emit one bare JSON object containing only direct SUPPORT, ATTACK, or REPHRASE edges to supplied IDs, or an empty relation list.

The adapter uses Qwen/Qwen3-0.6B with QLoRA r=16, alpha=32, three epochs, seed 20260823, and 2,048 private transformed QT30 training examples. It is the submission checkpoint because it had the best exact-patch result in the fixed v1 data-efficiency curve, while the subsequent hard-negative v2 experiment increased false edges and failed its preregistered material-improvement test. The training text is not redistributed. Consumers must obtain the official corpus and appropriate permission separately.

No tested training size cleared the preregistered reliability threshold. This checkpoint is published to make the assignment evidence reproducible, not as a claim that the behavior is reliable. The included manifests record the frozen eval/rubric hashes, deterministic metrics, data mix, fixed training config, adapter hash, and private-artifact hashes. The evaluation is small and corpus-specific; this is not evidence of general argument-mining reliability.

Load this PEFT adapter with a compatible Transformers/PEFT or Unsloth runtime. The training runtime used the 4-bit unsloth/qwen3-0.6b-unsloth-bnb-4bit mirror of the canonical Qwen/Qwen3-0.6B base.