add reproduce/install section

README.md

@@ -19,3 +19,22 @@ to an exact cell (with `table,row,col` coords). Unstated computational fields ar
 
 Reliability (grid-oracle over accuracy cells): **100% dataset-correct, 0 miss, ~82% full (dataset+model)**;
 schema-valid 100%; 0 fabricated numbers. See `result/case_studies/README.md`.
+
+## Reproduce
+The pipeline lives in the `parse_extract/` package of the code repo. Install + run:
+```bash
+# EXTRACT-only (no GPU) — re-run extraction on already-parsed papers:
+pip install openai==2.44.0 pyyaml==6.0.3
+
+# FULL parse+extract (CUDA-13 GPU):
+conda env create -f parse_extract/environment.yml
+conda activate paper2profile
+uv pip install -r parse_extract/requirements.txt --index-strategy unsafe-best-match --torch-backend=cu130
+
+# start the extract LLM server, then run the stage:
+GPU=0 bash rightsizing/engine/serve_qwen.sh &                                   # Qwen3-14B on :8001
+GPU=1 bash parse_extract/run.sh --config parse_extract/configs/full.yaml \
+      --steps parse,extract,verify,assemble                                     # -> data/extracted/<arxiv>/*.csv
+```
+Models: `opendatalab/MinerU2.5-Pro-2605-1.2B` (parse VLM) · `Qwen/Qwen3-14B` (extract structure-ID).
+Verified installable + runnable from a clean environment.