Read a prediction
Inspect the saved answer, evidence, lineage, warnings and execution state
A prediction is a durable run record, not just its generated output. Open it from the project Predictions list or use prediction_get with predictionId. Check execution and outcome separately before using the answer.
What to inspect
| Record information | Why it matters |
|---|---|
| Subject, horizon and horizonEndsAt | Establishes what and when the forecast addresses |
| Forecaster/version and query release | Identifies executable instructions and evidence logic |
| Evidence rows, diagnostics and snapshot/hash | Shows what supported the run and any retrieval limits |
| Prompt snapshot and model | Records the actual instructions and selected runtime model |
| Output, probability and confidence | Separates the answer from the model's support assessment |
| Warnings, error kind and attempt | Shows limitations or execution failure |
| Tokens, latency and cost status | Makes work and accounting availability visible |
| Evaluation versions and outcome | Connects the prediction to later observations |
Equipment reading example
Suppose a P-101 run outputs probability 0.7 and confidence 0.4. Those values are illustrative, not an expected answer to the fixture. The forecast suggests a 70% chance of the defined event, with weak support. Inspect whether the evidence actually contains P-101's measurements, their timestamp, service history coverage and relevant diagnostics.
The evidence snapshot may indicate hasMore. The forecasting loader keeps at most 100 rows; limited evidence can be incomplete even when execution completes. A prompt may request a limitation but the model can omit it; inspect the stored retrieval metadata yourself.
Reproducibility limits
The query release freezes logic, not every external source value. The invocation cutoff is recorded but does not automatically enforce historical source filtering. Replaying a prompt against today's live records can answer a different question. Keep actual evidence artifacts when comparing historical forecasts.
A confidence value is not a calibration report. An apparently precise probability or polished rationale is not an observed failure. Prediction output must be checked against evidence, then evaluated against real outcomes.
Follow up
Completed forecasts can be evaluated when the window permits. Failed or Cancelled runs do not have a completed typed forecast to score. Pending outcomes need observation, not another interpretation of the generated output. Use a new run for a new window or revised configuration; do not rewrite an older result to fit what happened.