agentic-evaluationlisted

# Agentic Evaluation ## Purpose and boundary Evaluate and improve Agent outputs through a bounded, evidence-based loop: ```text Generate → Evaluate → Critique → Refine → Verify ``` This is an evaluation workflow, not permission to run arbitrary tools, expose private data, or accept a model's self-rating as proof. Prefer deterministic checks and independent evaluators when available. Keep prompts, outputs, traces, and test data within the authorized audience; redact secrets and personal data before storing or sending them to a judge. ## Evaluation contract Define before generating or judging: - **task and expected outcome:** what the output must accomplish and what counts as failure; - **dimensions and weights:** accuracy, completeness, safety, clarity, format, latency, cost, or other relevant measures; - **evidence source:** deterministic test, reference answer, human review, retrieval citation, judge, or a combination; - **threshold and stop rule:** minimum acceptable score, maximum iterations, time/cost budget, and what happens when evidence conflicts; - **scope and permissions:** models, tools, datasets, environments, retention, and who may approve a release; - **report contract:** scores, failed dimensions, trajectory, uncertainty, limitations, and follow-up owner. Separate observed test results from judge interpretation and from the final decision. A score is not evidence of correctness unless its measurement method and coverage are stated. ## Step 1: establish