Prediction
Invoke a project forecaster for a subject and horizon
A Prediction node uses an existing project forecaster and its pinned published evidence-query release. It stores prediction output with the relevant model/evidence metadata. It is not a substitute for a configured query, model runtime or evaluation policy.
Before using it
Create or select the real resources described above in the pipeline’s project/workspace, inspect the upstream data shape and ensure your actor can perform this operation. A configuration fragment cannot create those resources. Read First pipeline for a complete graph and fixture.
Fields and bindings
| Field | Meaning |
|---|---|
forecasterId / forecasterSlug | Select an existing forecaster |
subjectRef | Subject reference matching the forecaster/query contract |
horizon / horizonEndsAt | Optional horizon label and required timestamp |
params | Parameters required by the forecaster/query |
Configure
Prepare the prediction node for the selected test resources. Use the operation and fields shown in this example. Show actual resource bindings, upstream/downstream ports and any write/model effects before applying the draft.Choose the forecaster, inspect the pinned evidence release and set a real subject, horizon end and required parameters. Keep the graph connected to a relevant ingress path for Studio validation; incoming rows do not automatically become query evidence.
Put this section under predict on a full node of kind predict. It is not a standalone API or MCP request. Replace resource placeholders, and provide the full node ports/policy and graph edges.
{
"forecasterId": "<FORECASTER_UUID>",
"subjectRef": "order-1",
"horizon": "review window",
"horizonEndsAt": "2027-01-01T00:00:00Z",
"params": {}
}Verify a bounded run
Inspect the stored prediction, schemas, source evidence/cutoff, warnings, confidence or probability and model usage. Evaluate against observed outcomes at the relevant time.
Validate, save a version and run a small known fixture. Inspect the actual node result and downstream consumer, not only the graph preview.
Limits and failure behavior
Forecasting must be enabled. The node is not automatically a per-row forecaster loop. Current v1 execution requires a literal subject string (or a literal expression); field-derived subject references are rejected. Success is not accuracy, and outcome recording does not automatically retrain a model.
The step returns predictionId, status: queued and nonBlocking: true. The pipeline does not wait for prediction completion; inspect that separate prediction job before consuming its output.