Forecast contracts
Reference forecaster settings, invocation, output and evaluation records
These contracts are available through the forecasting MCP tools and platform UI. Use the UI or assistant tabs in the task guides for authoring; the serialized definitions below describe technical fields, not a replacement for the product's forms.
Forecaster definition
| Field | Meaning |
|---|---|
| slug, displayName, description | Project identifier and human explanation |
| forecasterKind | Free-form subject-area label, such as equipment_failure |
| forecastKind | probability, categorical, numeric or scenario; controls scoring |
| queryDefinitionId | Published project evidence query; creation selects its latest published release |
| promptTemplate | Requires subject_ref, horizon and evidence_json variables |
| inputSchema, outputSchema | Draft-07 object contracts |
| executionConfig | single_call, chain or agentic with optional budget |
| defaultHorizon, defaultParams | Defaults merged at invocation |
| modelLabel | Optional provider/model identifier; configured deployment default if omitted |
| confidencePath, probabilityPath | Dot-separated output object paths |
| outcomeResolution, evaluationPolicy | Outcome/evaluation settings; keep consistent when authoring |
| status | active, disabled or archived |
A partial update preserves omitted fields. queryReleaseId on forecaster_update may be latest or a specific published release UUID. It is the evidence pin selector, not the query asset ID. Changing the evidence query selects its latest published release. Query evaluation policies have their own release pin.
Download a complete equipment forecaster payload. Replace the evidence query placeholder and add a unique idempotencyKey for forecaster_create. These examples use a project-scoped connection. On a workspace MCP endpoint, add projectId from project_list; use the schema discovered on your endpoint. Workspace scope comes from authentication.
Invocation
forecast_run requires forecasterId or forecasterSlug, subjectRef, a future horizonEndsAt and idempotencyKey. horizon may use the forecaster default. params default to an empty object and merge over forecaster defaults. Optional sourcePipelineRunId/watchlistId are lineage references, not evidence replacements.
Watchlist item
Requires forecasterId, subjectRef, schedule, evaluationHorizon and idempotencyKey. Optional horizon, params, enabled and nextRunAt configure timing. Use hourly/daily/weekly interval labels; arbitrary cron strings are not parsed by the current scheduler. evaluationHorizon is an interval such as 7 days, distinct from the prompt's horizon label.
Evaluation
prediction_evaluation_create requires predictionId, outcomeState, observedAt and idempotencyKey. Optional observedValue, evidenceAssertionIds, evaluatorKind and notes add detail. Supported evaluator kinds are human, query and external. Merely choosing external does not call a remote evaluator.
The prediction must be completed and typed. Observation timing and project assertion-reference checks apply. History is versioned; an evaluation changes the prediction's outcome status while preserving its generated output.
Records and limits
Prediction records retain version/release, horizon end, schemas, execution/evaluation snapshots, evidence cutoff/hash/artifact, result artifact, prompts, model, probability/confidence, usage, warnings and execution/outcome states. Inspect actual data availability; legacy/incomplete records may contain null metadata.
Evidence retrieval is bounded to 100 retained rows. Budgets limit configured execution, not the scope of permissions. Cost status may be unavailable. Read execution budgets and evaluation metrics for practical limits.