CODE EVALUATION PACKAGE
=======================

This package contains the frozen v5.2-rc5k-final DXF evaluator used in the CODE experiments, its full local dependency chain, reference-side semantic annotation records, and a portable batch-scoring/reporting entry point. Numerical metric modules under eval/ are unchanged from the source server; PROVENANCE.json records their hashes.

INSTALLATION
Use Python 3.12 and install requirements.txt:
  python -m pip install -r requirements.txt

DATA PREPARATION
Keep this directory under CODE/evaluation/. Extract CODE/reference_dxf.zip into CODE/, yielding CODE/reference_dxf/<id>.dxf. The evaluation data/ directory already contains the required compressed GT sidecars and population lists. Input images are not used by the scorer.

Place generated DXFs in a separate directory, named <sample_id>.dxf. Do not put reference files in the prediction directory except when deliberately running an identity sanity check.

RUN FROM CODE/evaluation/
  python evaluate.py --pred-dir /path/to/predicted_dxfs --population test1045 --output /path/to/new_results

To use the same common population as the paper's main comparison:
  python evaluate.py --pred-dir /path/to/predicted_dxfs --population paper1041 --output /path/to/new_paper_results

For validation, use --population validation254. If this package is elsewhere, pass --dataset-root /path/to/CODE or --gt-dir /path/to/reference_dxf. --ids-file accepts one ID per line for debugging; the output is explicitly marked as a custom subset and is not a complete-benchmark result.

POPULATIONS AND METRICS
- test1045: all 1,045 valid-reference test cases; CAMR denominator 915.
- paper1041: the frozen 1,041-case common comparison, also excluding four historical unavailable GPT requests; CAMR denominator 911. This historical exclusion is not a universal benchmark filter for new models.
- validation254: the established 254-case valid-reference validation set.

The output columns are SDR, SV, VC, GF, DF, AQ, LC, CAMR, CVR, LAR. Six quality scores use fixed-population 0--100 means with failed outputs and optional unavailable components zero-filled. CAMR uses only GT-eligible annotation cases and verified atomic matching evidence. CVR uses V=1, and LAR uses L>=0.85. No new Overall score is formed.

The paper's auxiliary ablation-quality table used a fixed intersection of 768 cases successfully delivered by all four variants. It is NOT the default population or each model's separate successful subset. To reproduce that table, select the same explicitly frozen intersection for every variant; do not compare separately filtered means.

EXECUTION STATUS AND SCOPE
This entry point scores already-generated DXF artifacts. It does not call a model, execute generated Python, reproduce generation token caps/timeouts, or provide a security sandbox. Run candidate programs separately in an appropriately isolated environment.

For paper-protocol evaluation, preserve the original generation/execution status in --failure-manifest, a JSONL containing {"id": "0000_00006012", "status": "execution_error"}, for example. Recognized statuses: scored, empty, syntax_error, execution_error, length_zero, safety_error, resource_error, missing_dxf, invalid_dxf, generation_failed. A recorded model failure contributes zero even if a partial DXF exists. Resource_error must describe a verified model resource failure, not broken infrastructure. API-unavailable and infrastructure failures must be resolved or handled by an explicitly shared population; they must not be silently relabeled as model failures.

Without a failure manifest, missing DXFs are treated as failed outputs, unreadable DXFs as invalid, and readable files are scored. Thus SDR is an artifact-delivery rate under the supplied population, not proof that generation constraints or execution security were reproduced. Unexpected scorer errors, missing references and GT-hash mismatches stop the run rather than becoming model zeros.

OUTPUTS
- summary.json and summary.csv: ten metrics, counts and denominator details.
- per_item.jsonl: per-sample scores, status and CAMR evidence state.
- raw/<id>.json: complete original evaluator output for each scored sample.
An existing output directory is never overwritten.

TESTS
  python -m unittest discover -s eval/tests -p "test_*.py"
  python -m unittest discover -s tests -p "test_*.py"

DATA PROVENANCE
Compressed sidecars are preserved byte-for-byte after decompression. Their original server paths remain historical provenance only. The CLI verifies each local GT against the frozen SHA-256 and substitutes its local path in memory; metric targets, view assignments, geometry, masks and thresholds are not rebuilt or altered.

No training loops, model weights, API credentials, server configuration, or private model outputs are included. Source evaluator modules retain internal full-GT and prompt-scoped functions because the final evaluator reuses them; evaluate.py reports the offline full-GT components only.