Data interface / 数据接口

Published dataset / 发布数据： images + paired metadata; train/test are separate. Labels and reference annotations must not enter the agent prompt. 图片配套元数据；训练集与测试集分开。标签和参考答案不进入 Agent 输入。

Agent input / 推理输入： runtime_input/cases.jsonl, one object per line:

{"case_id":"sample-001","image_path":"images/sample-001.jpg","image_sha256":"<64-character SHA-256>"}

Only these three keys are allowed. Image paths are relative to runtime_input/. The package also needs manifest.json; use python -m scripts.prepare_runtime_input --help to build a local package from metadata containing case_id and image_path. Other metadata is discarded. 仅保留这三个字段；转换工具不复制标签或私有标注，也不覆盖已有输出。

Training / 训练： SFT consumes reviewed trajectories, not image labels alone. PSD consumes frozen student prefixes, completion tokens, teacher top-k targets, and bound media; hints are teacher-only. SFT 使用审核后的轨迹；PSD 使用冻结前缀、续写 token、教师 top-k 分布及绑定图片。 Judge annotations stay separate. See training settings.