Single model call

Single-call execution loads the forecaster's pinned evidence query, renders the configured prompt, calls the selected model once and validates the final object against the output schema. Start here when one evidence set is sufficient; additional stages do not automatically improve accuracy.

Equipment configuration

You need an active project forecaster with a published equipment evidence query, a prompt defining the failure event and window, schemas requiring probability/confidence/rationale, and probabilityPath $.probability. Forecasting and the model/workflow runtime must be enabled. The complete equipment walkthrough includes the source and query if these resources do not yet exist.

{
  "mode": "single_call",
  "budget": {
    "maxDurationMs": 60000,
    "maxTotalTokens": 20000,
    "maxExternalCalls": 0
  }
}

This is an executionConfig object within the forecaster, not a complete definition. Zero external calls still allows the initial evidence query and model call; it disallows the optional external HTTP/plugin stages counted by this budget.

Configure and run

In the simple creation flow, single_call is the derived default. In Edit → Execution config, enter the object above. A connected assistant can propose forecaster_update with executionConfig and a unique idempotencyKey. After saving, inspect the version and evidence pin.

Open Run prediction, select this forecaster, use subject P-101, horizon 7d, a future evaluation date seven days after invocation and matching params. Through MCP use forecast_run with those values. Read prediction_get until it reaches a terminal state. An invocation queues work; it does not return a completed prediction immediately.

Inspect evidence and output

Check that only P-101 records reached the query, the measurement time is meaningful and missing history is acknowledged. Inspect the evidence artifact, rendered prompt and final output. Completed means execution and validation succeeded. It is separate from outcome status, which may remain Pending until observations are recorded.

The output's confidence is model-generated support, not a measured reliability percentage. Evaluate the probability against actual occurrence/non-occurrence under the same event definition and window.