Skip to main content
Save the run ID when you submit work. It is your route back to execution and its available evidence; a screenshot of a score is not a substitute for that record.

​
Know what happened

Read a result in this order:
  1. What did I request? Check the exact task set, actor and scoring method.
  2. What happened? Read the state and any failure or cancellation reason.
  3. What did it produce? Open the output and recorded assessment.
  4. What supports the result? Inspect the trajectory, verification output and available artifacts.
Replace <run-id> with the full ID returned by submission or synth run list. The same readers work for execution, evaluation and training.

​
Finished does not mean correct

A failed run may still have useful outputs or saved checkpoints. Inspect the evidence before discarding the work or submitting a replacement.

​
Counts have units and scope

For example, 15 batches × 4 tasks × 16 trials = 960 planned trials. That does not imply 960 successful assessments, 15 updates, or 60 distinct tasks. Dataset membership and the task-selection policy determine coverage. A recovery can create more physical trial records without creating a new experiment. Check the scope printed by an evidence page: an attempt-local trial count is not whole-run progress. An input checkpoint from an older run is not a checkpoint created by the new run. For training, inspect the actual model outputs:
Copy <checkpoint-id> from the checkpoint list; it is different from the run ID.

​
Read scores with their population

An average is meaningful only with its task set, scoring rule and coverage. If 17 of 18 trials have scores, an average over those 17 is not a fully scored 18-trial evaluation. The remaining trial might still be running, or it might have finished without an assessment; check which. Different task verifiers can use different scales. Do not assume that averaging all available scores weights every task equally or makes different datasets comparable. A training run changes the model over time; its per-group rewards are diagnostics, not a single fixed-model performance claim. For code review, examine valid findings and false positives. More findings do not necessarily make a reviewer more useful.

​
Follow work and preserve evidence

Ending a local event watch does not cancel the remote run. Check synth run show <run-id> again if the watch ends before execution does. Once the run has reached a terminal state, export its evidence:
Substitute the same run ID in the command and directory name. Use a new export directory and inspect the export’s coverage; retained evidence is not necessarily a complete historical sandbox image. For scripts, use --json and the documented command exit behavior. Automation explains how to avoid parsing human tables. Use Lifecycle and recovery and Errors and retry safety before retrying uncertain work.