# Which judgment survives a Label Studio YOLO export?

Checked 18 November 2026. When two active box annotations for the same image reached the Label Studio SDK 2.1.2 YOLO converter, the default export retained the older annotation's class. Reversing the input annotation array left the result unchanged. Swapping the two creation times changed the surviving class. The experiment used one fictional task, one local image path and two valid rectangles.

The converter sorts active annotations from newest to oldest by creation time and yields both. For the same image, the default YOLO writer builds the same label path for both annotations. In this fixture the older judgment's rectangle is written second and replaces the first file. The newer class is still listed in `classes.txt` because both classes are declared in the project schema. The Label Studio 1.23.2 release tag specifies SDK 2.1.2, and the generic YOLO conversion call leaves `split_labelers` at its default. [App dependency](https://github.com/HumanSignal/label-studio/blob/df96d79183156e556d91825d15fa7721f1145e1a/pyproject.toml#L75) · [App conversion call](https://github.com/HumanSignal/label-studio/blob/df96d79183156e556d91825d15fa7721f1145e1a/label_studio/data_export/models.py#L145-L170) · [SDK dispatch](https://github.com/HumanSignal/label-studio-sdk/blob/c20c899aaaa302740dacd46a47e5774bb677561f/src/label_studio_sdk/converter/converter.py#L294-L305) · [Creation-time sort](https://github.com/HumanSignal/label-studio-sdk/blob/c20c899aaaa302740dacd46a47e5774bb677561f/src/label_studio_sdk/converter/converter.py#L593-L626) · [Path and blank-result handling](https://github.com/HumanSignal/label-studio-sdk/blob/c20c899aaaa302740dacd46a47e5774bb677561f/src/label_studio_sdk/converter/converter.py#L1074-L1109) · [File replacement](https://github.com/HumanSignal/label-studio-sdk/blob/c20c899aaaa302740dacd46a47e5774bb677561f/src/label_studio_sdk/converter/exports/yolo.py#L153-L171)

| Case | Result |
| --- | --- |
| A older than B, either input array order | `sample.txt` contains A |
| B older than A | `sample.txt` contains B |
| Separate labeler folders, two distinct actor IDs | Actor 11's file contains A and actor 12's contains B |
| Separate labeler folders, same actor ID | Actor 11's file contains A |
| Older result empty and newer result B | `sample.txt` contains B |

The empty older result only creates an empty file if one does not exist. That branch limits the age claim to the two valid, nonempty judgments in the first three default runs. Separate labeler folders are an SDK option and do not preserve multiple judgments written by the same actor to the same path. The fixture does not establish how a project chose the input judgments or how any model would perform. [Option and default](https://github.com/HumanSignal/label-studio-sdk/blob/c20c899aaaa302740dacd46a47e5774bb677561f/src/label_studio_sdk/converter/converter.py#L966-L994)

For a delivery that is supposed to preserve review history, check a known disputed item across the source export and the delivered format. The final text file alone leaves the choice of judgment unexplained. If a team deliberately wants one judgment per image, record the selection or adjudication rule before flattening to delivery labels.

Run `python3 check.py | cmp - observations.json` with pipeline failure checking enabled. The standard-library [checker](check.py) hashes 92,423 bytes across four pinned SDK files. It executes the actual single-task file reader, annotation projection, rectangle conversion and label writer with a fixed schema and an existing fictional local path. The exact [observations](observations.json) record the input and processing order, the class file and all label paths and lines. Both runs produced byte-identical receipts. The four source files were also byte-matched against the PyPI 2.1.2 wheel, 1,066,859 bytes, SHA-256 `8b95a4021f0bb4bd5a4cdce38d998cfbc0028d68681f0fd3e9907525967013c3`, inspected in memory without installation. The experiment does not run the full service, its serializer or the SDK's task-array parser. [PyPI distribution](https://pypi.org/project/label-studio-sdk/2.1.2/#files)

An earlier report for Label Studio 1.11.0 describes the older judgment surviving after a new annotation was added to the same task. This SDK 2.1.2 fixture reproduces that retention effect on one image path with a fixed label vocabulary. The controls separate creation-time order from input order, actor grouping and an empty result. Separate reports cover duplicate basenames across distinct images and ambiguity between empty and unreviewed COCO annotations. The package contains original Python and invented observations. The checker fetches pinned public Apache-2.0 SDK source at runtime. [Earlier YOLO report](https://github.com/HumanSignal/label-studio/issues/5421) · [Duplicate basename report](https://github.com/HumanSignal/label-studio/issues/7096) · [Empty COCO report](https://github.com/HumanSignal/label-studio/issues/9774) · [SDK license](https://github.com/HumanSignal/label-studio-sdk/blob/c20c899aaaa302740dacd46a47e5774bb677561f/LICENSE)
