HORIZON / RELIABILITY LAB
Show what the agent actually did.
Evaluate tool execution, recovery, completion, and budgets against explicit fixtures. Inspect each failure before trusting an aggregate score.
Files are evaluated in this browser. This page does not call a model or upload your run data.
No evaluation loaded. Start with the example or import execution events from your harness.
What this measures
Exact fixture output, successful required tools, forbidden actions, unresolved calls, duplicate operations, ordered execution, retry recovery, and declared latency and cost budgets. A completion message alone does not pass.
These deterministic checks complement Horizon’s offline workflow evaluations. They do not establish general agent quality, semantic correctness beyond the fixture, or performance against another provider. Select fixtures before running the agent and retain failed runs.