ClawBenchAI Agent Benchmark
A repeatable test suite for comparing autonomous agents on quality, robustness, cost, and safety.
Eval Harness
The orchestration layer that runs tasks, captures outputs, scores results, and stores artifacts.
Prompt Injection
Malicious instruction patterns that attempt to override policy or exfiltrate hidden context.
Deterministic Replay
Re-running an evaluation with fixed conditions to verify a result can be reproduced.
Benchmark Drift
When benchmark tasks stop reflecting real-world workloads and scores become less meaningful.
Utility Retention
How much useful task performance remains after applying security defenses.