# Choosing a trial version

[中文](versions.zh-CN.md) · [Exercise](task.en.md)

| Version | Use | How to interpret it |
|---|---|---|
| 0.5.0 | Current multi-instance exercise | One dormant-member recall template, two task states, three fixed data instances including baseline. Each instance can run in either task state. Languages and repeated runs do not add business scenarios. |
| 0.4.1 | The original fixed trial archive | Keep the supplied archive unchanged. It contains the original two task states on the baseline data. The additional-instance step does not apply. |
| 0.3.0 | Historical model evidence | Historical episodes retain their original code, model and runtime metadata. They are not scores for 0.5.0 or its additional data instances. |

The original 0.4.1 archive SHA256 is `363069087914ceeb451cef38e862d968eb25bc66f2e68f36cb7ae5cfe86bf739`. Check the supplied digest before using an archive. A newer source checkout does not silently upgrade a downloaded archive; extract different versions into separate directories and use new result directories.

For a new multi-instance exercise, use 0.5.0 and the [current quick start](../quickstart.en.md). For a session already arranged around the fixed 0.4.1 package, use its own bundled instructions and record that version in the feedback. Do not mix files between archives.

The policy remains `prototype-v0.1`: consent, marketing contact frequency, refund treatment, approval and budget rules retain their meaning. A 1000-cent coupon is issued per eligible member, with a 3000-cent total ceiling; the budget need not be fully spent. Maintainer expectations are defined independently of the evaluated agent and are not agent inputs.

Current examples are deterministic scripts. They do not add model results, demonstrate production safety, or replace expert review. Identity-merging and refund-aggregation scenario ideas remain design proposals.