# DualXrayBench evaluation protocol

## Fixed evaluation set

Evaluate all 1,594 records in `annotations/benchmark.jsonl` in their published order. Supply both images referenced by `images.top` and `images.side` together with the unmodified `question`. Do not expose the `answer` or `target_instances` fields to the model, and do not train, tune prompts, or select checkpoints on benchmark answers.

## Multiple-choice scoring

Require a final choice from `A`, `B`, `C`, or `D`. Normalize only an unambiguous final choice; missing, multiple, or out-of-range choices are incorrect. Report exact accuracy as `correct / evaluated` for every task in `tasks.json`, micro accuracy over all 1,594 records, and the unweighted macro average of the eight task accuracies. Do not drop failed samples.

## Spatial evidence

`target_instances` preserves the paper-v1 reference object boxes as `[x_min, y_min, x_max, y_max]`. It supports spatial-error analysis and paper-compatible localization evaluation when a method emits boxes. A localization report must state its matching rule, IoU threshold, aggregation, and invalid-box handling; do not mix it with multiple-choice accuracy.

## Reproducibility

Report the exact model/revision, decoding parameters, prompt template, image preprocessing, and software version. Verify `checksums.sha256` before evaluation. Results must identify this package as `DualXrayBench v1.0.0` and include all eight per-task sample counts.