Real systems. Observable repairs.

Where does the agent break?

Inspect the attempts behind the score. Keep the model, its execution settings,
and the benchmark conditions visible as separate things.

A repair counts only when it survives
verification → restart → reboot.

Benchmark cohort

Provider

Agent harness

Reasoning

Scenario

Find a model

Select model rows to compare. Selected view ignores all filters except scenario.

· select a cell to inspect

✓Repaired×Failed–Unavailable— Not run

Whole-cohort summary · includes all scenarios, independent of the scenario filter
Configuration	Repairs	95% interval	Median time	Reported tokens	Known spend	Cost / repair	Cost coverage