Deploy and version an evaluation

​

Deploy it

Select deploy <key>@<version> at the bottom of the authoring page. The version is immutable once published: from then on, every session that finishes, and that its condition applies to, is scored by it.

​

See what is live

Analyze → eval authoring lists your organization’s hosted definitions, the evaluations the managed evaluator runs for it. Each row shows:

• its name, key, version, and result type
• its source checksum, which tells deployed revisions apart without opening the code
• whether it is conditional or runs on all completed sessions — the condition is what scopes an evaluation to particular agents or environments
• its timeout, its labels, and when it last changed

Search the list, or filter it by state. Evaluations your own worker registers are not listed here; their results carry a customer tag on the evaluations page, and hosted ones carry managed. An organization can have up to 100 different hosted evaluations enabled at once.

​

Publish a new version

Select new version on a row. The authoring page opens with that version’s code; change it, test it, and deploy it. Its key and result type carry over and cannot change. Publishing a successor disables its predecessor and keeps it on the list. Results keep the version that produced them, so a chart shows exactly when the new logic took over.

​

Roll back

Select disable on the current version and enable on the one you want back. Nothing is deleted, and every result stays as it was.

​

Stop an evaluation

Select disable. With no enabled version, it stops running on new sessions. To stop an evaluation your own worker runs, stop registering it: remove it from the worker, or stop the worker.

​

Score sessions you already have

Evaluation runs forward: a version deployed now never scores a session that ended before it. To score history, open score sessions you already have on the eval authoring page, choose a window of up to 90 days and, optionally, a single evaluation, and count before you run. The count is exactly what will run, and each session-and-evaluation pair in it is a billable evaluation. It fills gaps only. A session that already has a result for that evaluation keeps it, and running the same window twice scores nothing new. To score one session again — after a fix, or for a session that never ended cleanly — select re-evaluate on its page. The new result is added to the session’s history; earlier ones stay.

​

Permissions