Reproduces Table 3: mean absolute error (MAE) of four mechanical properties for each of the five models, computed over the 100 manually mapped row pairs.

table3_hea_cca/
├── hea_lib.py                # shared logic (readers, mapping parsers, MAE)
├── run_all.py                # runs all five models
├── table_1.csv               # ORIGINAL ground truth (shared by all models)
├── gpt54/  gpt52/  claude/  glm/  qwen/
│     ├── evaluate.py             # model-specific config + run
│     ├── <reconstructed>.csv     # RECONSTRUCTED dataset for this model
│     ├── row_mapping_100*.txt    # frozen 100-pair mapping used in the paper
│     └── results_reproduced.csv  # written when you run evaluate.py

python3 run_all.py            # all five models
python3 claude/evaluate.py    # one model

The frozen mapping file lists 100 (original_line, reconstructed_line) pairs. For each pair the four properties are read from table_1.csv and from the model's reconstructed CSV, and

MAE(property) = mean( |orig - recon| )  over pairs where BOTH values are numeric.

The reconstructed CSVs differ in column headers and formatting, and the original mapping generators parsed cells slightly differently. To reproduce the published numbers exactly, each evaluate.py selects the reader that matches its generator (all provided by hea_lib.py):

• claude — columns by header name, first-number cell parsing, skipinitialspace=False.
• glm / qwen — columns by header name, first-number parsing, skipinitialspace=True (their phase column contains quoted commas, which would otherwise shift later columns).
• gpt54 — fixed forward indices (6,7,8,9); safe_float (a cell that is not a clean number, e.g. "512 (compression)", is treated as missing) for sigma_y/sigma_max/epsilon; first-number parse for the E column (handles ground-truth entries like "135 (203)").
• gpt52 — ground truth by forward indices (6,7,8,9); reconstructed file by indices counted from the end of each row (its comma-containing phase column splits some rows, so the four property columns are read as the last four fields). Codex first-number parser throughout.

Line numbers are 1-based with the header on line 1 (first data row = line 2), matching how the mappings were generated.