Prediction kinds
Choose a probability, category, number or scenario and define how it will be evaluated
A prediction kind determines the meaning of the output and the metrics available later. Choose it before writing the prompt. In this section, equipment failure means an unplanned mechanical failure causing at least one hour of downtime within a stated window. Preventive maintenance is not that event.
Probability
Return a number from 0 to 1 for a binary event. For P-101 over seven days, probability 0.7 means a forecast of 70% likelihood, not a statement that a failure has already occurred. Configure forecastKind probability and probabilityPath $.probability. The output schema must require that numeric field with minimum 0 and maximum 1.
Confidence is a separate assessment of support or uncertainty, such as $.confidence. It is not the probability being scored. A model may return high likelihood and low confidence when history is incomplete.
Category
Define an unambiguous category and the observation that establishes it. For example, “maintenance priority at the evaluation date” could use an agreed high/medium/low label protocol. Configure forecastKind categorical and evaluationPolicy.forecastPath $.category when the output uses category. Merely recording that a failure occurred does not establish an observed maintenance-priority category.
Number
Define units, aggregation and window. “Unplanned downtime hours over seven days” is a numeric target; “downtime” alone is not. Configure forecastKind numeric and evaluationPolicy.forecastPath $.value for an output value. The evaluator must supply a numeric observedValue in the same units.
Scenario
Use scenario for a structured narrative about what might happen. Include conditions, evidence and limitations in the instructions. An outcome can be recorded, but the current scorer does not assign an automatic numeric scenario-quality metric. A plausible paragraph is not a measured prediction success.
Match schema and evaluation path
| Forecast kind | Minimal output fields | Evaluation settings | Actual observation |
|---|---|---|---|
| probability | probability (number 0–1), confidence | probabilityPath $.probability | occurred or not_occurred for the defined event |
| categorical | category (string), confidence | forecastPath $.category | observedValue matching your category vocabulary |
| numeric | value (number), confidence | forecastPath $.value | numeric observedValue in the declared units |
| scenario | scenario (string), confidence | Manual policy | Outcome and explanatory notes |
The simple creation form derives output fields from the selected kind. For category/number scoring, edit Outcome resolution to include the matching forecastPath and confirm the saved version. Without that path, the scorer's fallback is $.prediction, which does not match the form's category/value fields.
Create a kind-specific definition
You need a Semogram account with project access, a published evidence query for the relevant equipment and forecasting/model execution enabled. In the creation assistant, request the exact event or quantity, kind and window. Review the draft. In the manual form choose What does it predict?, Evidence and Horizon. After creation use Edit to align the schema and evaluation policy. A connected assistant can use forecaster_create with the complete contract, then forecaster_get to inspect the saved settings.
Changing the kind creates a new executable version. Keep comparisons grouped by compatible event, units, horizon and version; a downtime number cannot be compared directly with failure probability.