QLoRA (bf16 LoRA via PEFT+TRL) fine-tune of HuggingFaceTB/SmolLM3-3B for opinionated research triage: read-now / skim / skip / archive + novelty score + one-line reason + tags. Trained on a Sonnet-4.6-distilled dataset (true distillation). Judged blind vs its own base by a 3-lens panel (decisiveness / specificity / calibration): FT 9 / base 1 — beats base.

Training data: hqt2yotoz/verdict-engine-triage-v2 — 468 train + 52 val examples, Sonnet-4.6-distilled (true distillation).

The base over-praises (read-now + inflated novelty); the fine-tune discriminates (skim + honest novelty). github.com/setuc/verdict-engine