Overview
Produce evidence-backed forecasts and keep a track record against real outcomes
A forecaster defines a repeatable question about the future. A prediction is one recorded answer for a subject and a window. An evaluation records what actually happened and, where supported, scores that answer. These are separate records: successful execution does not establish that the prediction was accurate.
Semogram supports probabilities, categories, numbers and scenarios. A forecaster combines a published ontology evidence query, instructions, input/output schemas, a selected model, execution settings and an outcome policy. It belongs to a project in a workspace. The evidence query reads through that project's ontology bindings and workspace data endpoints.
Three records to understand
| Record | Purpose | Equipment example |
|---|---|---|
| Forecaster and immutable version | Defines the question, evidence release, model and output contract | Equipment failure risk, version 1 |
| Prediction | Saves a run's subject, window, evidence, output and execution history | Failure probability for pump P-101 over the next seven days |
| Evaluation | Records an observed outcome with time, evidence references and metrics | A breakdown occurred; an evaluation calculates Brier score |
A new prediction uses the active forecaster's current executable version. A saved version pins a specific published evidence-query release. Publishing another query release does not silently change that pin. Source values behind that release may still change.
Decide what the prediction means
Define the event or quantity before configuring a model. “Pump risk” is vague. “Unplanned mechanical failure causing at least one hour of downtime within seven days” defines what counts, the subject and the window. Keep probability separate from confidence: a likely event can still have weak supporting evidence.
| Kind | Example | Supported evaluation |
|---|---|---|
| Probability | Chance of a defined failure within seven days | Brier score and aggregate calibration |
| Categorical | Which maintenance priority band applies at the evaluation date | Exact-match accuracy against the observed category |
| Numeric | Total unplanned downtime hours over the window | Absolute and squared error |
| Scenario | A reasoned description of likely failure conditions | Recorded outcome and notes; no built-in numeric scenario score |
These are model-generated forecasts, not an automatic time-series training service. Recorded outcomes create a track record; they do not automatically retrain the selected model or rewrite prior predictions.
Prepare, run and evaluate
- Make the evidence query return relevant records for one known subject. Inspect age, coverage, identity and access before publishing it.
- Define the output and what observation will count as its outcome. Start with a bounded single model call.
- Review and create a forecaster. Inspect its immutable version and evidence release.
- Run once with a subject, horizon and explicit future evaluation date. Inspect the completed prediction's evidence and output.
- Record real observations when the window permits. Monitor evaluations and compare performance before changing versions.
The equipment example includes its own source fixture, model, bindings, query, forecaster and run instructions. The synthetic fixture verifies connectivity and contracts; it cannot demonstrate forecasting quality.
Choose the next task
| Task | Page |
|---|---|
| Choose probability, category, number or scenario | Prediction kinds |
| Connect and constrain evidence | Evidence query |
| Design instructions and output | Prompts and schemas |
| Add stages or bounded tool use | Chains and agentic execution |
| Schedule equipment subjects | Scheduling |
| Monitor execution and resolve operational failures | Run and operate |
| Record observations and measure performance | Outcomes and metrics |
Forecasting must be enabled on your deployment, with its model and workflow runtime configured. Account access, project permissions and evidence access still apply. Forecasting does not grant access to external systems or create missing data.