---
pretty_name: sortletters parallel 121 v21 OpenPI engineering fixture
tags:
- robotics
- lerobot
- engineering-fixture
- rejected-for-behavior-training
---

# sortletters_parallel_121_v21_openpi

> **STOP — rejected for behavior training.** This public revision is preserved only as a
> reproducible engineering fixture for loader, padding-mask, and training-throughput tests.
> It is not a valid real-robot policy-training dataset and must not be used to claim
> sortletters behavior quality or real-robot success rate.

> **警告：该数据已被行为训练资格门拒绝。** 仅允许用于 loader、mask 和训练吞吐工程验证；
> 不得用于声称有效真机策略训练、字母排序行为质量或真机成功率。

## Qualification evidence

Independent recomputation from the 121 `meta/retime_source_indices/episode_*.npz` files found:

- 65,532 / 66,053 frames (99.211%) use a right-front observation from a different source
  timestamp than the reference arm timeline.
- 65,533 frames compose left/right arm sources from different timestamps.
- 121 / 121 episodes contain cross-time composition.
- Median left/right source-time delta: 531 frames (17.7 seconds at 30 FPS).
- Maximum source-time delta: 868 frames.
- No jointly re-executed world observation or native success receipt repairs this causal mismatch.

The structured audit has SHA-256
`2a60f082001f977b141fd718634e9c4932201c7bd92414411e390c1494a1d9d3` and status
`rejected_for_behavior_training`.

## Provenance

- Derived dataset source repo: `Shiki42/sortletters_parallel_121`
- Source revision: `76fd84d2f254f7acc95d5a5b4caafc00a03207a1`
- Initial immutable payload transport commit: `78783c2e34087cb8d099f022f75c877aa7afc363`
- Payload: 489 files, 2,026,345,313 bytes

The data files at the initial payload commit remain content-addressed and unchanged. Later commits
only add or revise this warning card and do not upgrade behavior-training qualification.

## Allowed use

- loader unit tests
- `action_is_pad` and 12D-to-32D loss-mask tests
- training-runtime and throughput engineering

## Prohibited claims

- valid real-robot policy training
- sortletters behavior quality
- real-robot success rate