Add zero-shot baseline mode: --adapter_path none skips adapter loading, --no_think suppresses Qwen3 thinking"

Add evaluate_v3.py — stratified sampling, layer-aware max tokens, incremental saves, resume support

Fix: flush stdout for nohup, log every sample, add timestamps

Add evaluate_v2.py — standard-aware KPI checking (fixes 92% false negatives in reliability metric)

Fix: --max_seq_length → --max_length (matches train.py argparse)

Add comprehensive scientific documentation with 38 citations

Upload README.md

fix: max_seq_length → max_length, warmup_ratio → warmup_steps (TRL 1.3 compat)

Upload run.sh

Upload evaluate.py

Upload train.py

Upload requirements.txt

initial commit