NavidBroumandfar/agent-behavior-evals-lab

Agent Behavior Evals Lab is a local evaluation harness for testing AI assistants and agentic systems against policy-defined behavior expectations. It’s intended for engineers, adapter developers, evaluators, and researchers who need reproducible, policy-driven assessments of agent behavior (including future adapters and saved transcripts). It is not an OpenClaw-only project and the current milestone uses a deterministic mock client. Key features: a behavior policy and failure taxonomy; 30 JSONL eval cases across categories; target profiles and system prompts; a deterministic mock model client and rule-based scorer; end-to-end mock eval runner producing scored JSONL traces and a Markdown baseline report. Tooling includes schema/manifest validators, trace/report writers, saved-transcript replay, adapter fixture import/validation, adjudication-aware reporting, regression/comparison checks, and utilities for promoting reviewed outputs.