This folder contains everything needed to reproduce the two quantitative benchmark tables in the paper directly from the released data:

• Table 2 — MeltpoolNet benchmark (precision / recall / F1 of selected fields + melting-temperature MAE), in table2_meltpoolnet/
• Table 3 — High-entropy / complex-concentrated alloy (HEA/CCA) benchmark (MAE of four mechanical properties), in table3_hea_cca/

Each table is evaluated for the same five backbone models:

Each model's score is computed against a frozen manual row mapping — the 100 ground-truth ↔ extracted row pairs that were mapped by hand for the paper (see the paper's Materials and Methods). For each table/model we ship:

• the original ground-truth dataset (shared per table),
• the model's reconstructed/extracted dataset,
• the row-mapping file that pairs 100 ground-truth rows with 100 extracted rows, and
• an evaluate.py that re-reads those three artifacts and recomputes the reported numbers (writing a results_reproduced.csv next to itself).

The evaluation does not re-run any LLM or re-do the mapping — it deterministically recomputes the published metrics from the released data and the prior mapping.

• MeltpoolNet (Table 2): 100 row pairs mapped on five fields — material composition, laser power, scan velocity, layer thickness, and paper DOI.
• HEA/CCA (Table 3): 100 row pairs mapped on three fields — alloy composition, source paper number, and Vickers hardness (HV).

• Python 3.8+ — standard library only, no third-party packages.

# Table 2 — all five models
cd table2_meltpoolnet && python3 run_all.py

# Table 3 — all five models
cd ../table3_hea_cca && python3 run_all.py

Or run a single cell, e.g. the Claude column of Table 3:

cd table3_hea_cca/claude && python3 evaluate.py

See each table's own README.md for the per-cell expected values and the exact matching rules.

All twenty table cells (10 model/table combinations across both tables) reproduce the published numbers exactly.