sortletters_serialized_121

STOP — not qualified as a coherent behavior-training baseline. The files are internally indexed and decodable, but the available provenance does not establish a valid real-robot sortletters policy-training dataset or success rate.

警告：该数据未通过一致行为训练基线资格门。 文件可读取、索引一致，但不能据此声称 有效真机字母排序策略或真机成功率。

Read-only source audit findings for revision 344ef8e12b2f9f5e5bc4c25be042dde085cabb12:

• No persisted per-episode success labels or autonomous success-verifier records exist.
• Episodes 84–89 contain visible human presence and direct object intervention during the task.
• All 121 exact terminal right-front frames visually show M I N T, while the only persisted prompt is 字母排序; it does not specify MINT and naturally conflicts with alphabetical I M N T.
• All 121 exact start frames use one fixed letter permutation; no letter-permutation diversity was observed.
• Wrist videos contain repeated decoded frames during associated-arm motion, and original per-sensor capture timestamps were not recorded.

At most 115 episodes remain a candidate for a separately named, fixed-layout MINT-rearrangement dataset after explicit semantic relabeling and additional trajectory review. This card does not certify those episodes as successes.

Allowed use without further qualification: dataset inspection, loader tests, and runtime engineering.