Documentation / guides
Cases and runs
Move from validated evaluation cases to bounded runs and durable results.
The cases-and-runs section follows one evaluation from its durable definition
to its recorded result. Start with a Scenario and its optional expectation,
group scenarios into a revisioned Suite, then let eval.Run produce a report
of observations and assessments.
Choose the contract
| Need | Continue with | What you will learn |
|---|---|---|
| Define input, identity, revision, and expectations | Cases and suites | How validation keeps fixtures stable before a target runs. |
| Execute trials with cancellation and bounded workers | Runs and results | How target errors, evaluator errors, and quality outcomes remain distinct. |
| Persist a safe canonical report | Reporting | How report/v1 and FileSink preserve evidence without raw payloads. |
Keep the suite revision and target subject revision aligned. A mismatch is a target-stage error, so a run cannot quietly evaluate evidence under the wrong model or agent identity.
Source
The case contract is defined in
scenario.go and
suite.go, and the
execution contract is in
run.go.
Proof
Validation behavior is exercised by
scenario_test.go,
and run-stage behavior by
run_test.go.