Evaluation

Metrics first, interpretation second. No arbitrary LLM scores.

Simple explanation

If you cannot fail in a visible way, you are not evaluating — you are storytelling.

Technical explanation

Task completion, accuracy, cost, latency, hallucinations, tool-use. Keep versions. Later SignalLab: early-signal precision, not a vibe check.

Evidence

Profiling A/B/C (profile-llm.v1) and numbered HITL automation (hitl-agent.v2: classification 100%, false autonomy 15%, down from 60% in v1). Versioned gold, not a framework ranking.