Versions and evidence pins
Keep prior predictions reproducible when instructions, models or queries change
A forecaster is a project record with immutable executable versions. Creation publishes its first version. An update publishes another version; there is no separate forecaster_publish MCP action. Runs keep the version and evidence-query release used at invocation.
What a version preserves
The version snapshots the evidence release, forecast kind, prompt, input/output schemas, execution configuration, evaluation policy, model provider/model identifier, confidence/probability paths and a snapshot hash. A prediction records those identities and execution snapshots alongside its evidence and result.
When the evidence pin changes
| Edit | Evidence release used by the new version |
|---|---|
| Change prompt, model, schema or budget | Existing pin is retained |
| Publish another release of the same query | Existing forecaster remains pinned |
| Set queryReleaseId to latest | Latest published release is selected |
| Set queryReleaseId to a published release UUID | That release is selected, if it belongs to the query |
| Select a different evidence query | Latest published release of the new query |
In the edit UI, inspect Evidence release and choose deliberately. For an equipment forecaster, a connected assistant can call:
{
"jsonrpc": "2.0",
"id": 1,
"method": "tools/call",
"params": {
"name": "forecaster_update",
"arguments": {
"forecasterId": "<EQUIPMENT_FORECASTER_UUID>",
"queryReleaseId": "latest",
"idempotencyKey": "<UNIQUE_REPIN_KEY>"
}
}
}This is a partial update: omitted fields retain stored values. Inspect forecaster_get afterward to verify the actual pin and current version. A retired selected release cannot be published into a new forecaster version.
Roll out a change
Check the revised query on known equipment, publish its release, then create a forecaster version with an explicit pin. Run the same representative subjects under comparable windows and inputs. Keep older completed prediction records rather than overwriting their output.
Version pins stabilize definitions; they do not guarantee identical model generation or immutable live data. Evidence snapshots, observation time, model availability and source changes still matter.
Pause and retire
Only Active forecasters run. Disabled forecasters are paused; Archived ones are retired from new work. Pausing does not establish an outcome for pending predictions. Inspect watchlists and existing runs separately. Use the UI management action or forecaster_archive when retiring the record, subject to the relevant permissions and confirmation.