active · 2026
CLEAR-S Eval Harness
A layered evaluation system for testing agent outputs, trajectories, safety, latency, and cost across prompts, frameworks, and models.
- MLflow
- agents
- evaluation
- traces
Most agent evaluations inspect the final answer. CLEAR-S treats the execution path as part of the product.
The harness combines deterministic checks, trace-level invariants, model-based judges, safety tests, latency, and cost into seven practical axes: correctness, latency, execution, adherence, relevance, safety, and cost.
The question it answers
Did the system improve, or did one metric move while a hidden failure got worse?
Each experiment records the configuration, dataset slice, outputs, traces, and evaluator results. This makes it possible to compare a prompt change, a workflow fix, a held-out domain, or a model swap using the same language.
Why trajectory checks matter
A policy assistant may cite a valid document without verifying that the document supports its claim. An output check sees a citation. A trajectory check sees whether verification happened before citation.
The harness is designed to make those rules explicit, cheap to run, and easy to inspect.