Operating predictions
Manage scheduled forecasts, execution failures and unresolved outcomes
Operating predictions includes launching due subjects, checking model execution and checking what actually happened. Those are separate stages with separate records. A Completed prediction can remain Pending for its outcome until its window closes or sufficient evidence establishes an early occurrence.
You need a Semogram account with project access and the permissions for the action. Reading predictions uses project read access; creating/updating watchlist items and evaluations uses project write access. Forecast invocation uses queries:execute; cancellation uses runs:execute. Deployment operators also maintain the authorized background jobs.
Equipment example
An Active Equipment failure risk forecaster predicts whether P-101 will have an unplanned mechanical failure causing at least one hour downtime in the next seven days. Its published evidence query selects that equipment. A daily watchlist creates repeated predictions; each invocation has its own seven-day window and pinned forecaster/evidence versions.
At 08:00 on Monday, prediction A covers Monday through the following Monday. Tuesday's prediction B has a separate window. One real incident can be relevant to both windows, but each evaluation must reference the appropriate prediction and establish its event definition. Do not count the forecasts as independent incidents.
Three kinds of background work
| Work | Purpose | What to inspect |
|---|---|---|
| Watchlist scheduling | Create predictions for due enabled subjects | nextRunAt, lastRunAt, forecaster state and new prediction IDs |
| Queued delivery recovery | Re-deliver eligible stale queued prediction work | Existing prediction/workflow identity and dispatch diagnostics |
| Outcome resolution | Execute a pinned outcome policy for due Pending predictions | Verdict evidence, immutable evaluation and diagnostic notes |
These jobs must be configured on the deployment. A watchlist save is not a scheduler. Re-delivery is not a general retry of a failed model or external write. Outcome resolution does not retrain the model.
Operating guides
- Scheduling: cadence, evaluation intervals, pausing and versions.
- Monitoring: execution, evidence completeness, usage and unresolved outcomes.
- Failures and recovery: uncertain requests, cancellation and deliberate reruns.
- Outcome operations: due windows, manual review, query failures and corrections.
The first prediction guide supplies a complete test setup. Synthetic equipment readings verify connectivity and execution; use real observations and a suitable evaluation cohort before making predictive-quality claims.