Read your first run

Goal

One run fully read and judged: what it did, what it produced, whether to trust it — and a recorded decision (approve, retry, or correct). Done means no finished run in your workspace is a mystery.

Image: run view with impact summary leading and evidence available behind it.

Steps

1. Open runs and history and pick a finished run — ideally the pipeline you committed in Run your first pipeline.
2. Start with the impact summary, not the mechanics. What changed, what was produced, what needs a human? The timeline and node states exist for investigation, not for orientation.
3. Read the AI narrative of the run next: what the system believes happened, in order. Compare it against what you expected the workflow to do. Mismatches between narrative and expectation are the cheapest defects to catch.
4. Follow one output into its evidence: source records, transforms, and steps. Then check freshness — when the inputs were synced and whether anything upstream changed since.
5. If something failed, use diagnostics before acting: which node, what error class (permissions, unreachable source, bad shape, policy block), and whether it is transient or structural. The category decides the fix, not the message text.

Image: diagnostics pointing at one failed node with the error class and suggested fix.

6. Choose exactly one next action:
   • Approve — outputs are right; record it so future runs learn.
   • Retry — the failure was transient (source hiccup, timeout) and nothing about the plan needs changing.
   • Correct — the plan, mapping, scope, or data is wrong. Say what, fix it at its layer, and rerun. Never silently rerun a structural failure hoping for a different result.
7. Record the decision on the run. Unreviewed runs are invisible debt; reviewed runs are training data.

When runs confuse

• Green but wrong: the plan executed; the intent was mis-scoped. Correct the request or the mapping, not the run.
• Fails the same way twice: stop retrying. It is structural — read the diagnostics category and fix the layer it points at.
• Blocked on approval: that is the system working. Review the gated node in permissions view and approve or reject deliberately.

What good looks like

• Every finished run has a recorded review decision.
• Retries are rare and justified; corrections land at the right layer.
• You can explain any run's outcome from its evidence alone.

Next

• Make review systematic: Operators.
• When outputs feed decisions: Get your first forecast.
• Something broken: Troubleshooting.